Origins of Answer-Set Programming - Some Background And Two Personal Accounts

08/16/2011 ∙ by Victor W. Marek, et al. ∙ 0

We discuss the evolution of aspects of nonmonotonic reasoning towards the computational paradigm of answer-set programming (ASP). We give a general overview of the roots of ASP and follow up with the personal perspective on research developments that helped verbalize the main principles of ASP and differentiated it from the classical logic programming.



There are no comments yet.


page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction — Answer-Set Programming Now

Merely ten years since the term was first used and its meaning formally elaborated, answer-set programming has reached the status of a household name, at least in the logic programming and knowledge representation communities. In this paper, we present our personal perspective on influences and ideas — most of which can be traced back to research in knowledge representation, especially nonmonotonic reasoning, logic programming with negation, constraint satisfaction and satisfiability testing — that led to the two papers Marek and Truszczyński (1999); Niemelä (1999) marking the beginning of answer-set programming as a computational paradigm.

Answer-set programming (ASP, for short) is a paradigm for declarative programming aimed at solving search problems and their optimization variants. Speaking informally, in ASP a search problem is modeled as a theory in some language of logic. This representation is designed so that once appended with an encoding of a particular instance of the problem, it results in a theory whose models, under the semantics of the formalism, correspond to solutions to the problem for this instance. The paradigm was first formulated in these terms by Marek and Truszczyński (1999) and Niemelä (1999).

The ASP paradigm is most widely used with the formalism of logic programming without function symbols, with programs interpreted by the stable-model semantics introduced by Gelfond and Lifschitz (1988). Sometimes the syntax of programs is extended with the strong negation operator and disjunctions of literals are allowed in the heads of program rules. The semantics for such programs was also defined by Gelfond and Lifschitz (1991). They proposed to use the term answer sets for sets of literals, by which programs in the extended syntax were to be interpreted. Ten years after the answer-set semantics was introduced, answer sets lent their name to the budding paradigm. However, there is more to answer-set programming than logic programming with the stable-model and answer-set semantics. Answer-set programming languages rooted directly in first-order logic, extending it in some simple intuitive ways to model definitions, have also been proposed over the years and have just matured to be computationally competitive with the original logic programming embodiments of the paradigm (Denecker, 1998; Denecker and Ternovska, 2008; East and Truszczyński, 2006).

Unlike Prolog-like logic programming, ASP is fully declarative. Neither the order of rules in a program nor the order of literals in rules have any effect on the semantics and only negligible (if any) effect on the computation. All ASP formalisms come with the functionality to model definitions and, most importantly, inductive definitions, in intuitive and concise ways. Further, there is a growing body of works that start addressing methods of modular program design (Dao-Tran, Eiter, Fink, and Krennwallner, 2009; Janhunen, Oikarinen, Tompits, and Woltran, 2009) and program development and debugging (Brain and Vos, 2005; Brummayer and Järvisalo, 2010). These features facilitate modeling problems in ASP, and make ASP an approach accessible to non-experts.

Most importantly, though, ASP comes with fast software for processing answer-set programs. Processing of programs in ASP is most often done in two steps. The first step consists of grounding the program to its equivalent propositional version. In the second step, this propositional program is solved by a backtracking search algorithm that finds one or more of its answer sets (they represent solutions) or determines that no answer sets (solutions) exist. The current software tools employed in each step, commonly referred to as grounders and solvers, respectively, have already reached the level of performance that makes it possible to use them successfully with programs arising from problems of practical importance.

This effectiveness of answer-set programming tools is a result of a long, sustained and systematic effort of a large segment of the Knowledge Representation community, and can be attributed to a handful of crucial ideas, some of them creatively adapted to ASP from other fields. Specifically, domain restriction was essential to help control the size of ground programs. It was implemented in lparse, the first ASP grounder Niemelä and Simons (1996). The well-founded semantics Van Gelder et al. (1991) inspired strong propagation methods implemented in the first full-fledged ASP solver smodels Niemelä and Simons (1996). Program completion Clark (1978) provided a bridge to satisfiability testing. For the class of tight programs Erdem and Lifschitz (2003), it allowed for a direct use of satisfiability testing software in ASP, the idea first implemented in an early version of the solver cmodels111 Loop formulas Lin and Zhao (2002) extended the connection to satisfiability testing to arbitrary programs. They gave rise to such successful ASP solvers as assat Lin and Zhao (2002), pbmodels Liu and Truszczyński (2005) and later implementations of cmodels Lierler and Maratea (2004). Database techniques for query optimization influenced the design of the grounder for the dlv (Leone, Pfeifer, Faber, Eiter, Gottlob, Perri, and Scarcello, 2006). Important advances of satisfiability testing including the data structure of watched literals, restarts, and conflict-clause learning were incorporated into the ASP solver, at present the front-runner among ASP solvers and the winner of one track of the 2009 SAT competition. Some of the credit for the advent of high-performance ASP tools is due to the initiative to hold ASP grounder and solver contests. The two editions of the contest so far (Gebser, Liu, Namasivayam, Neumann, Schaub, and Truszczyński, 2007; Denecker, Vennekens, Bond, Gebser, and Truszczynski, 2009) focused on modeling and on solver performance, and introduced a necessary competitive element into the process.

The modeling features of ASP and computational performance of ASP software find the most important reflection in a growing range of successful applications of ASP. They include molecular biology (Gebser, Guziolowski, Ivanchev, Schaub, Siegel, Thiele, and Veber, 2010a, b), decision support system for space shuttle controllers (Balduccini, Gelfond, and Nogueira, 2006), phylogenetic systematics (Erdem, 2011), automated music composition (Boenn, Brain, Vos, and Fitch, 2011), product configuration (Soininen and Niemelä, 1998; Tiihonen, Soininen, Niemelä, and Sulonen, 2003; Finkel and O’Sullivan, 2011) and repair of web-service workflows (Friedrich, Fugini, Mussi, Pernici, and Tagni, 2010).

And so, ASP is now a declarative programming paradigm built on top of a solid theoretical foundation, with features that facilitate its use in modeling, with software supporting effective computation, and with a growing list of successful applications to its credit. How did it all come about? This paper is an attempt to reconstruct our personal journey to ASP.

2 Knowledge Representation Roots of Answer-Set Programming

One of the key questions for knowledge representation is how to model commonsense knowledge and how to automate commonsense reasoning. The question does not seem particularly relevant to ASP understood, as it now commonly is, as a general purpose computational paradigm for solving search problems. But in fact, knowledge representation research was essential. First, it recognized and emphasized the importance of principled modeling of commonsense and domain knowledge. The impact of the modeling aspect of knowledge representation and reasoning is distinctly visible in the current implementations of ASP. They support high level programming that separates modeling problem specifications from problem instances, provide intuitive means to model aggregates, and offer direct means to model defaults and inductive definitions. Second, knowledge representation research, and especially nonmonotonic reasoning research, provided the theoretical basis for ASP formalisms: the answer-set semantics of programs can be traced back to the semantics of default logic and autoepistemic logic, the semantics of the logic FO(ID) (Denecker, 2000; Denecker and Ternovska, 2008) has it roots in the well-founded semantics of nonmonotonic provability operators.

In this section we discuss the development of those ideas in knowledge representation that eventually took shape of answer-set programming. In their celebrated 1969 paper, McCarthy and Hayes wrote

[…] intelligence has two parts, which we shall call the epistemological and the heuristic. The epistemological part is the representation of the world in such a form that the solution of problems follows from the facts expressed in the representation. The heuristic part is the mechanism that on the basis of the information solves the problem and decides what to do.

With this paragraph McCarthy and Hayes ushered knowledge representation and reasoning into artificial intelligence and moved it to one of the most prominent positions in the field. Indeed, what they referred to as the

epistemological part is now understood as knowledge representation, while the heuristic part

has evolved into broadly understood automated reasoning — a search for proofs or models.

The question how to do knowledge representation and reasoning quickly reached the forefront of artificial intelligence research. McCarthy suggested first-order logic as the formalism for knowledge representation. The reasons behind the proposal were quite appealing. First-order logic is “descriptively universal” and proved itself as the formal language of mathematics. Moreover, key reasoning tasks in first-order logic could be automated, assuming one adopted appropriate restrictions to escape semi-decidability of first-order logic in its general form.

However, there is no free lunch and it turned out that first-order logic could not be just taken off the shelf and used for knowledge representation with no extra effort required. The problem is that domain knowledge is rarely complete. More often than not, information available to us has gaps. And the same is true for artificial agents we would like to function autonomously on our behalf. Reasoning with incomplete knowledge is inherently defeasible

. Depending on how the world turns out to be (or depending on how the gaps in our knowledge are closed), some conclusions reached earlier may have to be withdrawn. The monotonicity of first-order logic consequence relation is at odds with the

nonmonotonicity of defeasible reasoning and makes modeling defeasible reasoning in first-order logic difficult.

To be effective even when available information is incomplete, humans often develop and use defaults, that is, rules that typically work but in some exceptional situations should not be used. We are good at learning defaults and recognizing situations in which they should not be used. In everyday life, it is thanks to defaults that we are not bogged down in the qualification problem McCarthy (1977), that is, normally we do not check that every possible precondition for an action holds before we take it. And we naturally take advantage of the frame axiom McCarthy and Hayes (1969) when reasoning, that is, we take it that things remain as they are unless they are changed by an action. Moreover, we do so avoiding the difficulties posed by the ramification problem Finger (1987), which is concerned with side effects of actions. However, first-order logic conspicuously lacks defaults in its syntactic repertoire nor does it provide an obvious way to simulate them. It is not at all surprising, given that defaults have a defeasible flavor about them. Not being aware that a situation is “exceptional” one may apply a default but later be forced to withdraw the conclusion upon finding out the assumption of “non-exceptionality” was wrong.

Yet another problem for the use of first-order logic in knowledge representation comes from the need to model definitions, most notably the inductive ones. The way humans represent definitions has an aspect of defeasibility that is related to the closed-world assumption. Indeed, we often define a concept by specifying all its known instantiations. We understand such a definition as meaning also that nothing else is an instance of the concept, even though we rarely if ever say it explicitly. But the main problem with definitions lies elsewhere. Definitions often are inductive and their correct meaning is captured by the notion of a least fixpoint. First-order logic cannot express the notion of a least fixpoint and so does not provide a way to specify inductive definitions.

These problems did not go unrecognized and in late 1970s researchers were seeking ways to address them. Some proposals called for extensions of first-order logic by explicit means to model defaults while other argued that the language can stay the same but the semantics had to change. In 1980, the Artificial Intelligence Journal published a double issue dedicated to nonmonotonic reasoning, a form of reasoning based on but departing in major ways from that in first-order logic. The issue contained three papers by McCarthy (1980), Reiter (1980), and McDermott and Doyle (1980) that launched the field of nonmonotonic reasoning.

McCarthy’s proposal to bend the language of first-order logic to the needs of knowledge representation was to adjust the semantics of first-order logic and to base the entailment relation among sentences in first-order logic on minimal models only McCarthy (1980). He called the resulting formalism circumscription and demonstrated how circumscription could be used in several settings where first-order logic failed to work well. Reiter (1980) extended the syntax of first-order logic by defaults, inference rules with exceptions, and described formally reasoning with defaults. Reiter was predominantly interested in reasoning with normal defaults but his default logic was much more general. Finally, McDermott and Doyle proposed a logic based on the language of modal logic which, as they suggested, was also an attempt to model reasoning with defaults. This last paper was found to suffer from minor technical problems. Two years later, McDermott (1982) published another paper which corrected and extended the earlier one.

These three papers demonstrated that shortcomings of first-order logic in modeling incomplete knowledge and supporting reasoning from these representations could be addressed without giving up on the logic entirely but by adjusting it. They sparked a flurry of research activity directed at understanding and formalizing nonmonotonic reasoning. One of the most important and lasting outcomes of those efforts was the autoepistemic logic proposed by Moore (1984, 1985). Papers by Moore can be regarded as closing the first phase of the nonmonotonic reasoning as a field of study.

Identifying nonmonotonic reasoning as a phenomenon deserving an in-depth study was a major milestone in logic, philosophy and artificial intelligence. The prospect of understanding and automating reasoning with incomplete information, of the type we humans are so good at, excited these research communities and attracted many researchers to the field. Accordingly, the first 10-12 years of nonmonotonic reasoning research brought many fundamental results and established solid theoretical foundations for circumscription McCarthy (1980); Lifschitz (1988), default logic (Reiter and Criscuolo, 1981; Hanks and McDermott, 1986; Marek and Truszczyński, 1989; Pearl, 1990), autoepistemic logic (Moore, 1985; Niemelä, 1988; Marek and Truszczyński, 1991; Shvarts, 1990; Schwarz, 1991) and modal nonmonotonic logics in the style of McDermott and Doyle (Marek, Shvarts, and Truszczyński, 1993; Schwarz, 1992; Schwarz and Truszczyński, 1992). Researchers made progress in clarifying the relationship between these formalisms Konolige (1988, 1989); Marek and Truszczyński (1989b); Bidoit and Froidevaux (1991); Truszczyński (1991). Computational aspects received much attention, too. First complexity results appeared in late 1980s and early 1990s Cadoli and Lenzerini (1990); Marek and Truszczyński (1991); Kautz and Selman (1989); Gottlob (1992); Stillman (1992) and early, still naive at that time, implementations of automated reasoning with nonmonotonic logics were developed around the same time Etherington (1987); Niemelä and Tuominen (1986, 1987); Ginsberg (1989). Several research monographs were published in late 1980s and early 1990s systematizing that phase of nonmonotonic reasoning research and making it accessible to outside communities Besnard (1989); Brewka (1991); Marek and Truszczyński (1993).

Expectations brought up by the advent of nonmonotonic reasoning formalisms were high. It was thought that nonmonotonic logics would facilitate concise and elaboration tolerant representations of knowledge, and that through the use of defeasible inference rules like defaults it would support fast reasoning. However, around the time of the first Knowledge Representation and Reasoning Conference, KR 1989 in Toronto, concerns started to surface in discussions and papers.

First, there was the issue of multiple belief sets, depending on the logic used represented as extensions or expansions. A prevalent interpretation of the problem was that multiple belief sets provided the basis for skeptical and brave modes of reasoning. Skeptical reasoning meant considering as consequences a reasoner was sanctioned to draw only those formulas that were in every belief set. Brave reasoning required a non-deterministic commitment to one of the possible belief sets with all its elements becoming consequences of the underlying theory (in the nonmonotonic logic at hand). The first approach was easy to understand and accept at the intuitive level. But as a reasoning mechanism it was rather weak as in general it supported few non-trivial inferences. The second approach was underspecified — it provided no guidelines on how to select a belief set, and it was not at all obvious how if at all humans perform such a selection. Both skeptical and brave reasoning suffered from the fact that there were no practical problems lying around that could offer some direction as to how to proceed with any of these two approaches.

Second, none of the main nonmonotonic logics seemed to provide a good formalization of the notion of a default or of a defeasible consequence relation. This was quite a surprising and in the same time worrisome observation. Nonmonotonic reasoning brought attention to the concept of default and soon researchers raised the question of how to reason about defaults rather than with defaults (Pearl, 1990; Kraus, Lehmann, and Magidor, 1990; Lehmann and Magidor, 1992). A somewhat different version of the same question asked about defeasible consequence relations, whether they can be characterized in terms of intuitively acceptable axioms, and whether they have semantic characterizations Gabbay (1989); Makinson (1989). Despite the success of circumscription, default and autoepistemic logics in addressing several problems of knowledge representation, it was not clear if or how they could contribute to the questions above. In fact, it still remains an open problem whether any deep connection between these logics and the studies of abstract nonmonotonic inference relations exists.

Next, the complexity results obtained at about same time Marek and Truszczyński (1991); Eiter and Gottlob (1993a); Gottlob (1992); Stillman (1992); Eiter and Gottlob (1993b, 1995) were viewed as negative. They dispelled any hope of higher computational efficiency of nonmonotonic reasoning. Even under the restriction to the propositional case, basic reasoning tasks turned out to be as complex as and in some cases even more complex (assuming polynomial hierarchy does not collapse) than reasoning in propositional logic. Even more discouraging results were obtained for the general language.

Finally, the questions of applications and implementations was becoming more and more urgent. There were no practical artificial intelligence applications under development at that time that required nonmonotonic reasoning. Nonmonotonic logics continued to be extensively studied and discussed at AI and KR conferences, but the belief that they can have practical impact was diminishing. There was a growing feeling that they might amount to not much more but a theoretical exercise. Complexity results notwithstanding, the ultimate test of whether an approach is practical can only come from experiments, as the worst-case complexity is one thing but real life is another. But there was little work on implementations and one of the main reasons was lack of test cases whose hardness one could control. Researchers continued to analyze “by hand” small examples arguing about correctness of their default or autoepistemic logic representations. These toy examples were appropriate for the task of understanding basic reasoning patterns. But they were simply too easy to provide any meaningful insights into automated reasoning algorithms and their performance.

And so the early 1990s saw a growing sentiment that in order to prove itself, to make any lasting impact on the theory and practice of knowledge representation and, more generally, on artificial intelligence, practical and efficient systems for nonmonotonic reasoning had to be developed and their usefulness in a broad range of applications demonstrated. Despite of all the doom and gloom of that time, there were reasons for optimism, too. The theoretical understanding of nonmonotonic logics reached the level when development of sophisticated computational methods became possible. Complexity results were disappointing but the community recognized that they concerned the worst case setting only. Human experience tells us that there are good reasons to think that real life does not give rise to worst-case instances too often, in fact, that it rarely does. Thus, through experiments and the focus on reasoning with structured theories one could hope to obtain efficiency sufficient for practical applications. Moreover, it was highly likely that once implemented systems started showing up, they would excite the community, demonstrate the potential of nonmonotonic logics, and spawn competition which would result in improvements of algorithms and performance advances.

It is interesting to note that many of the objections and criticisms aimed at nonmonotonic reasoning were instrumental in helping to identify key aspects of answer-set programming. Default logic did not provide an acceptable formalization of reasoning about defaults but inspired the answer-set semantics of logic programs Bidoit and Froidevaux (1987); Gelfond and Lifschitz (1988, 1991) and helped to solve a long-standing problem of how to interpret negation in logic programming. Answer-set programming, which adopted the syntax of logic programs, as well as the answer-set semantics, can be regarded as an implementation of a significant fragment of default logic. The lack of obvious test cases for experimentation with implementations forced researchers to seek them outside of artificial intelligence and led them to the area of graph problems. This experience showed that the phenomenon of multiple belief sets can be turned from a bug to a feature, when researchers realized that it allows one to model arbitrary search problems, with extensions, expansions or answer sets, depending on the logic used, representing problem solutions (Cadoli, Eiter, and Gottlob, 1997; Marek and Remmel, 2003).

However important, knowledge representation was not the only source of inspiration for ASP. Influences of research in several other areas of computer science, such as databases, logic programming and satisfiability, are also easily identifiable and must be mentioned, if only briefly. One of the key themes in research in logic programming in the 1970s and 1980s was the quest for the meaning of the negation operator. Standard logic programming is built around the idea of a single intended Herbrand model. A program represents the declarative knowledge about the domain of a problem to solve. Some elements of the model, more accurately, ground terms the model determines, represent solutions to the problem. All works well for Horn programs, with the least Herbrand model of a Horn program as the natural choice for the intended model. But the negation operator, being ingrained in the way humans describe knowledge, cannot be avoided. The logic programming community recognized this and the negation was an element of Prolog, an implementation of logic programming, right from the very beginning. And so, the question arose for a declarative (as opposed to the procedural) account of its semantics.

Subsequent studies identified a non-classical nature of the negation operator. This nonmonotonic aspect of the negation operator in logic programming was also a complicating factor in the effort to find a single intended model of logic programs with negation. It became clear that to succeed one either had to restrict the class of programs or to move to the three-valued settings. The first line of research resulted in an important class of stratified programs (Apt, Blair, and Walker, 1988), the second one led Fitting (1985) and Kunen (1987) to the Kripke-Kleene model and, later on, Van Gelder, Ross, and Schlipf (1991) to the well-founded model.

In the hindsight, the connection to knowledge representation and nonmonotonic reasoning should have been quite evident. However, the knowledge representation and logic programming communities had little overlap at the time. And so it was not before the work by Bidoit and Froidevaux (1987) and Gelfond (1987) that the connection was made explicit and then exploited. That work demonstrated that intuitive constraints on an intended model cannot be reconciled with the requirement of its uniqueness. In other words, with negation in the syntax, we must accept the reality of multiple intended models. The connection between logic programming and knowledge representation, especially, default and autoepistemic logics was important. On the one hand, it showed that logic programming can provide syntax for an interesting non-trivial fragment of these logics, and drew attention of researchers attempting implementations of nonmonotonic reasoning systems. On the other hand, it led to the notion of a stable model of a logic program with negation. It also reinforced the importance of the key question how to adapt the phenomenon of multiple intended models for problems solving.

The work in databases provided a link between query languages and logic programming. One of the outcomes of this work was , a fragment of logic programming without function symbols, proposed as a query language. The database research resulted in important theoretical studies concerning complexity, expressive power and connection of to the SQL query language (Cadoli et al., 1997). was implemented, for instance as a part of DB2 database management system. introduced an important distinction between extensional and intentional database components. Extensional database is the collection of tables that are stored in the database, the corresponding relation names known as extensional predicate symbols. The intensional database is a collection of intentional tables defined by queries. In time this distinction was adopted by answer-set programming as a way to separate problem specification from data. The database community also considered extensions of with the negation connective. Because of the semantics of the resulting language, multiplicity of answers in was a problem, as it was in a more general setting of arbitrary programs with negation. Therefore, never turned into a practical database query language (although, its stratified version could very well be used to this end). However, it was certainly an interesting fragment of logic programming. And even though its expressive power was much lower than that of general programs,444It has to be noted though that the expressive power of general programs with function symbols and negation goes well beyond what could be accepted as computable under all reasonable semantics (Schlipf, 1995; Marek, Nerode, and Remmel, 1994). there was hope that fast tools to process can be developed. Jumping ahead, we note here that it was that was eventually adopted as the basic language of answer-set programming.

3 Towards Answer-Set Programming at the University of Kentucky

Having outlined some of the key ideas behind the emergence of answer-set programming, we now move on to a more personal account of research ideas that eventually resulted in the formulation of the answer-set programming paradigm. In this section, Victor Marek and Mirek Truszczynski, discuss the evolution of their understanding of nonmonotonic logics and how they could be used for computation that led to their paper Stable logic programming — an alternative logic programming paradigm Marek and Truszczyński (1999). A closely intertwined story of Ilkka Niemelä, follows in the subsequent section. As the two accounts are strongly personal and necessarily quite subjective, for the most part they are given in the first person. And so, in this section “we” and us refers to Victor and Mirek, just as “I” in the next one to Ilkka.

In mid 1980s, one of us, Victor, started to study nonmonotonic logics following a suggestion from Witold Lipski, his former Ph.D. student and close collaborator. Lipski drew Victor’s attention to Reiter’s papers on closed-world assumption and default logic Reiter (1978, 1980). In 1984, Victor attended the first Nonmonotonic Reasoning Workshop at Mohonk, NY, and came back convinced about the importance of problems that were discussed there. In the following year, he attracted Mirek to the program of the study of mathematical foundations of nonmonotonic reasoning.

In 1988 Michael Gelfond visited us in Lexington and in his presentation talked about the use of autoepistemic logic Moore (1985) to provide a semantics to logic programs. At the time we were already studying autoepistemic logic, inspired by talks Victor attended at Mohonk and by Moore’s paper on autoepistemic logic in the Artificial Intelligence Journal Moore (1985). We knew by then that stable sets of formulas of modal logic, introduced by Stalnaker (1980) and shown to be essential for autoepistemic logic, can be constructed by an iterated inductive definition from their modal-free part Marek (1989). We also realized the importance of a simple normal form for autoepistemic theories introduced by Konolige (1988).

Thus, we were excited to see that logic programs can be understood as some simple autoepistemic theories thanks to Gelfond’s interpretation Gelfond (1987). Soon thereafter, we also realized that logic programs could be interpreted also as default logic theories and that the meaning of logic programs induced on them by default logic extensions is the same as that induced by autoepistemic expansions Marek and Truszczyński (1989b). It is important to note that default logic was first used to assign the meaning to logic programs by Bidoit and Froidevaux (1987), but we did not know about their work at the time. Bidoit and Froidevaux effectively defined the stable model semantics for logic programs. They did so indirectly and with explicit references to default extensions. The direct definition of stable models in logic programming terms came about one year later in the celebrated paper by Gelfond and Lifschitz (1988).

What became apparent to us soon after Gelfond’s visit was that despite both autoepistemic expansions and default extensions inducing the same semantics on logic programs, it was just serendipidity and not the result of the inherent equivalence of the two logics. In fact, we noticed that there was a deep mismatch between Moore’s autoepistemic logic with the semantics of expansions and Reiter’s default logic with the semantics of extensions. In the same time, we discovered a form of default logic, to be more precise, an alternative semantics of default logic, which was the perfect match for that of expansions for autoepistemic logic Marek and Truszczyński (1989). This research culminated about 15 years later with a paper we co-authored with Marc Denecker that provided a definitive account of the relationship between default and autoepistemic logics (Denecker, Marek, and Truszczyński, 2003) and resolved problems and flaws of an earlier attempt at explaining the relationship due to Konolige (1988). Another paper in this volume (Denecker, Marek, and Truszczynski, 2011) discusses the informal basis for that work and summarizes all the key results.

The relationship between default and autoepistemic logic was of only marginal importance for the later emergence of answer-set programming. But another result inspired by Gelfond’s visit turned out to be essential. In our study of autoepistemic logic we wanted to establish the complexity of the existence of expansions. We obtained the result by showing that the problem of the existence of a stable model of a logic program is NP-complete and, by doing so, we obtained the same complexity for the problem of the existence of expansions of autoepistemic theories of some simple form but still rich enough to capture logic programs under Gelfond’s interpretation Marek and Truszczyński (1991).

The result for autoepistemic logic did not turn out to be particularly significant as the class of autoepistemic theories it pertained to was narrow. And it was soon supplanted by a general result due to Gottlob (1992), who proved the existence of the expansions problem to be -complete. But it was an entirely different matter with the complexity result concerning the existence of stable models of programs!

First, our proof reduced a combinatorial problem, that of the existence of a kernel in a directed graph, to the existence of stable model of a suitably defined program. This was a strong indication that stable semantics may, in principle, lead to a general purpose formalism for solving combinatorial and, more generally, search problems. Of course we did not fully realize it at the time. Second, it was quite clear to us, especially after the first KR conference in Toronto in May 1989 that the success of nonmonotonic logics can come only with implementations. Many participants of the conference (we recall David Poole and Matt Ginsberg being especially vocal) called for working systems. Since by then we understood the complexity of stable-model computation, we asked two University of Kentucky students Elizabeth and Eric Freeman to design and implement an algorithm to compute stable models of propositional programs. They succeeded albeit with limits — the implementation could process programs with about 20 variables only. Still, theirs was most likely the first working implementation of stable-model computation. Unfortunately, with the M.S. degrees under their belts, Eric and Elizabeth left the University of Kentucky.

For about three years after this first dab into implementing reasoning systems based on a nonmonotonic logic, our attention was focused on more theoretical studies and on the work on a monograph on mathematical foundations of nonmonotonic reasoning based on the paradigm of context-dependent reasoning. However, the matter of implementations had constantly been on the backs of our minds and in 1992 we decided to give the matter another try. As we felt we understood default logic well and as it was commonly viewed as the nonmonotonic logic of the future, in 1992 we started the project, Default Reasoning System DeReS. We aimed at implementing reasoning in the unrestricted language of propositional default logic. We also started a side project to DeReS, the TheoryBase project, aimed at developing a software system generating default theories to be used for testing DeReS. The time was right as two promising students, Pawel Cholewinski and Artur Mikitiuk, joined the University of Kentucky to pursue doctorate degrees in computer science.

As is common in such circumstances, we were looking for a sponsor of this research and found one in the US Army Research Office (US ARO), which was willing to support this work. A colleague of ours, Jurek Jaromczyk, also at the University of Kentucky, coined the term DeReS, a pun on an old polish word “deresz” presently rarely used and meaning a stallion, quite appropriate for the project to be conducted in Lexington, “the world capital of the horse.” In the proposal to US ARO we promised to investigate basic reasoning problems of default logic:

  1. Computing of extensions

  2. Skeptical reasoning with default theories — testing if a formula belongs to all extensions of an input default theory

  3. Brave reasoning with default theories — testing if a formula belongs to some extension of an input default theory.

The basic computational device was backtracking search for a basis of an extension of a finite default theory . This was based on two observations due to Reiter: that while default extensions of a finite default theory are infinite, they are finitely generated; and that the generators are all formulas of and the consequent formulas of some defaults from . We also employed ideas such as relaxed stratification of defaults Cholewiński (1995); Lifschitz and Turner (1994) for pruning the search space and relevance graphs for simplifying provability.

We also thought it was important to have the nonmonotonic reasoning community accept the challenge of developing implementations of automated nonmonotonic reasoning. Our proposal to US ARO contained a request for funding of a retreat dedicated to knowledge representation, nonmonotonic reasoning and logic programming. The key goals for the retreat were:

  1. To stimulate applications of nonmonotonic formalisms and implementations of automated reasoning systems based on nonmonotonic logics

  2. To promote the project to create a public domain library of benchmark problems in nonmonotonic reasoning.

We held the workshop in Shakertown, KY, in October 1994. Over 30 leading researchers in nonmonotonic reasoning participated in talks and we presented there early prototypes of DeReS and TheoryBase. Importantly, we heard then for the first time from Ilkka Niemelä about the work on systems to perform nonmonotonic reasoning in the language of logic programs in his group at the Helsinki University of Technology. The meeting helped to elevate the importance of implementations of nonmonotonic reasoning systems and their applications. It evidenced first advances in the area of implementations, as well as in the area of benchmarks, essential as so far most problems considered as benchmarks were toy problems such as “Tweety” and “Nixon Diamond.”

The DeReS system was not designed with any specific applications in mind. At the time we believed that, since default logic could model several aspects of commonsense reasoning, once DeReS became available, many artificial intelligence and knowledge representation researchers would use it in their work. And we simply regarded broadly understood knowledge representation problems as the main application area for DeReS.

Working on DeReS immediately brought up to our attention the question of testing and performance evaluation. In the summer of 1988, Mirek attended a meeting on combinatorics where Donald Knuth talked about the problem of testing graph algorithms and his proposal how to do it right. Knuth was of the opinion that testing algorithms on randomly generated graphs is insufficient and, in fact, often irrelevant. Graphs arising in real-life settings rarely resemble graphs generated at random from some probabilistic model. To address the problem, Knuth developed a software system, Stanford GraphBase, providing a mechanism for creating collections of graphs that could be then used in projects developing graph algorithms. Graphs produced by the Stanford GraphBase were mostly generated from real-life objects such as maps, dictionaries, novels and images. Some were based on rather obscure sources such as sporting events in Australia. The documentation was superb (the book by Knuth on the Stanford GraphBase is still available). The Stanford GraphBase was free and its use was not restricted. From our perspective, two aspects were essential. First, the Stanford GraphBase provided a unique identifier to every graph it created and so experiments could be described in a way allowing others to repeat them literally and perform comparisons on identical sets of graphs. Second, the Stanford GraphBase supported creating families of examples similar but increasing in size, thus allowing to test scalability of algorithms being developed.

In retrospect, the moment we started talking about testing our implementations of default logic was the defining moment on our path towards the answer-set programming paradigm. Based on our complexity result concerning the existence of stable models and its implication for default logic, we knew that all NP-complete graph problems could be reduced to the problem of the existence of extensions. The reductions expressed instances of graph problems as default theories. Thus, in order to get a family of default theories, similar but growing in size, we needed to select an NP-complete problem on graphs (say, the hamiltonian cycle problem), generate a family of graphs, and generate for each graph in the family the corresponding default theory. These theories could be used to test algorithms for computing extensions. This realization gave rise to the TheoryBase, a software system generating default theories based on reductions of graph problems to the existence of the extension problem and developed on top of the Stanford GraphBase, which served as the source of graphs. The TheoryBase provided default theories based on six well-known graph problems: the existence of

-colorings, Hamiltonian cycles, kernels, independent sets of size at least , and vertex covers of size at most . As the Stanford GraphBase provided an unlimited supply of graphs, the TheoryBase offered an unlimited supply of default theories.

We will recall here the TheoryBase encoding of the existence of a -coloring problem as it shows that already then some fundamental aspects of the methodology of representing search problems as default theories started to emerge. Let be an undirected graph with the set of vertices . Let be a set of colors. To express the property that vertex is colored with , we introduced propositional atoms . For each vertex , and for each color , we defined the default rule

The set of default rules models a constraint that vertex obtains exactly one color. The default theory , where

has extensions corresponding to all possible colorings (not necessarily proper) of the vertices of . Thus, the default theory defines the basic space of objects within which we need to search for solutions. In the present-day answer-set programming implementations choice or cardinality rules, which offer much more concise representations, are used for that purpose. Next, our TheoryBase encoding imposed constraints to eliminate those colorings that are not proper. To this end, we used additional default rules, which we called killing defaults, and which now are typically modeled by logic program rules with the empty head. To describe them we used a new propositional variable and defined

for each edge of the graph and for each color . Each default “kills” all color assignments which give color to both ends of edge . It is easy to check (and it also follows from now well-known more general results) that defaults of the form “kill” all non-proper colorings and leave precisely those that are proper. This two-step modeling methodology, in which we first define the space of objects that contains all solutions, and then impose constraints to weed away those that fail some problem specifications, constitutes the main way by which search problems are modeled in ASP.

The key lesson for us from the TheoryBase project was that combinatorial problems can be represented as default theories and that constructing these representations is easy. It was then for the first time that we sensed that programs finding extensions of default theories could be used as general purpose problem solving tools. It also lead us, in our internal discussions to thinking about “second-order” flavor of default logic, given the way it was used for computation. Indeed, in all theories we developed for the TheoryBase, extensions rather than their single elements represented solutions. In other words, the main reasoning task did not seem to be that of skeptical or brave reasoning (does a formula follow skeptically or bravely from a default theory) but computing entire extensions. We talked about this second-order flavor when presenting our paper on DeReS at the KR conference in 1996 (Cholewiński, Marek, and Truszczyński, 1996). At that time, we knew we were closing in on a new declarative problem-solving paradigm based on nonmonotonic logics.

A problem for us was, however, a fairly poor performance of DeReS. The default extensions are closed under consequence. This means that processing of default theories requires testing provability of prerequisites and justifications of defaults. This turned out to be a major problem affecting the processing time of our implementations. It is not surprising at all in view of the complexity results of Gottlob (1992) and Stillman (1992). Specifically, existence of extensions is a -complete problem.

There is, of course, an easy case of provability when all formulas in a default theory are conjunctions of literals only. Now the problem with the provability of premises disappears. However, DeReS organized its search for solutions by looking for sets of generating defaults, inheriting this approach from the case of general default theories, rather than for literals generating an extension. And that was still a problem. There are typically many more rules in a default theory than atoms in the language.

At the International Joint Conference and Symposium on Logic Programming in 1996, Ilkka and his student Patrik Simons presented the first report on their smodels system Niemelä and Simons (1996). But it seems fair to say that only a similar presentation and a demo Ilkka gave at the Logic Programming and Non-Monotonic Reasoning Conference in 1997, in Dagstuhl, made the community really take notice. The lparse/smodels constituted a major conceptual breakthrough and handled nicely all the traps DeReS did not avoid. First, lparse/smodels focused on the right fragment of default logic, logic programming with the stable-model semantics. Next, it organized search for a stable model by looking for atoms that form it. Finally, it supported programs with variables and separated, as was the standard in logic programming and databases, a program (a problem specification) from an extensional database (an instance of the problem).

The work by Niemelä had us focus our thinking about nonmonotonic logics as computational devices on the narrower but all-important case of logic programs. We formulated our ideas about the second-order flavor of problem solving with nonmonotonic logics and contrasted them with the traditional Prolog-style interpretation of logic programming. We stated our initial thoughts on the methodology of problem solving that exploited our ideas of modeling combinatorial problems that we used in the TheoryBase project, as well as the notion of program-data separation that came from the database community and was, as we just mentioned, already used in our field by Niemelä. These ideas formed the backbone of our paper on an alternative way logic programming could be used for solving search problems Marek and Truszczyński (1999).

4 Towards Answer-Set Programming at the Helsinki University of Technology

In this section Ilkka Niemelä discusses the developments at the Helsinki University of Technology that led to the paper Logic Programs with Stable Model Semantics as a Constraint Programming Paradigm Niemelä (1999). Similarly as in the previous section, the account is very personal and quite subjective. Hence, in this section ”I” refers to Ilkka.

I got exposed to nonmonotonic reasoning when I joined the group of Professor Leo Ojala at the Helsinki University of Technology in 1985. The group was studying specification and verification techniques of distributed systems. One of the themes was specification of distributed systems using modal, in particular, temporal and dynamic logics. The group had got interested in the solutions of the frame problem based on nonmonotonic logics when looking for compact and computationally efficient logic-based specification techniques for distributed systems. My role as a new research assistant in the group was to examine autoepistemic logic by Moore, nonmonotonic modal logics by McDermott and Doyle, and default logic by Reiter from this perspective.

There was a need for tool support and together with a doctoral student Heikki Tuominen we developed a system that we called the Helsinki Logic Machine, “an experimental reasoning system designed to provide assistance needed for application oriented research in logic” (Niemelä and Tuominen, 1986, 1987). The system included tools for theorem proving, model synthesis, model checking, formula manipulation for modal, temporal, epistemic, deontic, dynamic, and nonmonotonic logics. It was written in Quintus Prolog and contained implementations, for instance, for Reiter’s default logic, McDermott and Doyle style nonmonotonic modal logic, and autoepistemic logic in the propositional case based on the literature and some own work (Etherington, 1987; McDermott and Doyle, 1980; Niemelä, 1988). While nonmonotonic reasoning was a side-track in the Helsinki Logic Machine, it seems that it was one of the earliest working nonmonotonic reasoning systems although we were not very well aware of this at the time.

The work and, in particular, the difficulties in developing efficient tools led me to further investigations to gain a deeper understanding of algorithmic issues and related complexity questions (Niemelä, 1988, 1988, 1990, 1992). Similar questions were studied by others and in the early 90s results explaining the algorithmic difficulties started emerging. These results showed that key reasoning tasks in major nonmonotonic logics are complete for the second level of the polynomial hierarchy Cadoli and Lenzerini (1990); Gottlob (1992); Stillman (1992); Niemelä (1992). This indicated that these nonmonotonic logics have two orthogonal sources of complexity that we called classical reasoning and conflict resolution. Orthogonality means that even if we assume that classical reasoning can be done efficiently, nonmonotonic reasoning still remains NP-hard (unless the polynomial hierarchy collapses).

These results made me to focus more on conflict resolution to develop techniques for pruning the search space of potential expansions/extensions. One approach was to develop compact characterizations of expansions/extensions capturing their key ingredients. For autoepistemic logic I developed such a characterization based on the idea that expansions can be captured in terms of the modal subformulas in the premises and classical reasoning and exploited the idea in a decision procedure for autoepistemic logic Niemelä (1988). Together with Jussi Rintanen we also showed that if one limits the theory in such a way that conflict resolution is easy by requiring stratification, then efficient reasoning is possible by further restrictions affecting the other source of complexity Niemelä and Rintanen (1992).

The characterization based on modal subformulas generalizes also to default logic where extensions can be captured using justifications in the rules and leads to an interesting way of organizing the search for expansions/extensions as a binary search tree very similar to that in the DPLL algorithm for SAT Niemelä (1994, 1995). Further pruning techniques can be integrated to cut substantial parts of the potential search space for expansions/extensions and exploit, for instance, stratified parts of the rule set. My initial but very unsystematic experimentation gave promising results.

In 1994 encouraged and challenged by the Shakertown Workshop organized by Victor and Mirek, I decided to restrict to a simple subclass of default theories, that is, logic programs with the stable model semantics. For this subclass classical reasoning is essentially limited to Horn clauses and can be done efficiently in linear time using techniques proposed by Dowling and Gallier in the 1980s Dowling and Gallier (1984). I had no particular application in mind. The goal was to study whether the conflict resolution techniques I had developed for autoepistemic and default logic would scale up so that it would be possible to handle very large sets of rules which meant at that time thousands or even tens of thousands of rules.

At that time Patrik Simons joined my group and started working on a C++ implementation of the general algorithm tailored to logic programs. Patrik had excellent insights into the key implementation issues from very early on and the first version was released in 1995 Niemelä and Simons (1995). The C++ implementation was called smodels and it computed stable models for ground normal programs. It gave surprising good results immediately and could handle programs with a few thousand ground rules. Challenge benchmarks were combinatorial problems, mainly colorability and Hamiltonian cycles, an idea that I learnt from Mirek and Victor in Shakertown. For such hard problems the performance of smodels was substantially better than state-of-the-art tools such as the SLG system Chen and Warren (1996).

When developing benchmarks for evaluating novel algorithmic ideas and implementation techniques we soon realized that working with ground programs is too cumbersome. In practice, for producing large enough interesting ground programs for benchmarking we needed to write separate programs in some other language to generate ground logic programs. This took considerable time for each benchmark family and was quite inflexible and error-prone. We realized that in order to attract users and to be able to attack real applications we needed to support logic program rules with variables.

For handling rules with variables we decided to employ a two level architecture. The first phase was concerned with grounding, a process to generate a set of ground instances of the rules in the program so that stable models are preserved. Actual stable-model computation was taking place in the second model search phase on the program grounded in the previous one. The idea was to have a separation of concern, that is, be able to exploit advanced database and other such techniques in the first phase and novel search and pruning techniques in the other in such a way that both steps could be developed relatively independently. We released the first such system in 1996 Niemelä and Simons (1996).

This was a major step forward in attracting users and getting closer to applications. Such a system supporting rules with variables enabled compact and modular encodings of problems without any further host language. It was now also possible to separate the problem specification and the data providing the instance to be solved.

Working with the system and studying potential applications made me realize that logic programming with the stable model semantics is very different from traditional logic programming implemented in various Prolog systems. These systems are answering queries by SLD resolution and producing answer substitutions as results. But we were using logic programs more like in a constraint programming approach where rules are seen as constraints on a solution set (stable model) of the program and where a solution is not an answer substitution but a stable model, that is, a valuation that satisfies all the rules. This is like in constraint satisfaction problems where a solution is a variable assignment satisfying all the constraints. I wrote down these ideas in a paper Logic Programs with Stable Model Semantics as a Constraint Programming Paradigm which was first presented in a workshop on Computational Aspects of Nonmonotonic Reasoning in 1998 Niemelä (1998) and then appeared as an extended journal version in 1999 Niemelä (1999). The paper emphasized, in particular, the knowledge representation advantages of logic programs as a constraint satisfaction framework:

“Logic programming with the stable model semantics is put forward as an interesting constraint programming paradigm. It is shown that the paradigm embeds classical logical satisfiability but seems to provide a more expressive framework from a knowledge representation point of view.”

In 1998 we put more and more emphasis on potential applications and, in particular, on product configuration. This made us realize that a more efficient grounder supporting an extended modeling language is needed. At that point another student, Tommi Syrjänen, with excellent implementation skills and insight on language design, joined the group and work on a new grounder, lparse, started. The goal was to enforce a tighter typing of the variables in the rules to facilitate the application of more advanced database techniques for grounding and the integration of built-in predicates and functions, for instance, for arithmetic.

We also realized that for many applications normal logic programs were inadequate not allowing compact and intuitive encodings. This led to the introduction of new language constructs: (i) choice rules for encoding choices instead of recursive odd loops needed in normal programs and (ii) cardinality and weight constraints for typical conditions needed in many practical applications Soininen and Niemelä (1998); Niemelä et al. (1999). In order to fully exploit the extensions computationally Patrik Simons developed techniques to provide built-in support for them also in the model search phase in the version 2 of smodels Simons (1999).

So in 1999 when Vladimir Lifschitz coined the term answer-set programming, the system that we had with lparse as the grounder and smodels version 2 as the model search engine offered quite promising performance. For example, for propositional satisfiability the performance of smodels compared nicely to the best SAT solvers at that time (before more efficient conflict driven clause learning solvers like zchaff emerged). Moreover, very interesting serious application work started. For example, at the Helsinki University of Technology we cooperated with the product data management group on automated product configuration which eventually led to a spin-off company Variantum ( Moreover, in Vienna the dlv project for handling disjunctive programs had started a couple years earlier and had already made promising progress.

5 Conclusions

Now, more than 12 years since ASP became a recognizable paradigm of search problem solving, we see that the efforts of researchers in various domains: artificial intelligence, knowledge representation, nonmonotonic reasoning, satisfiability and others resulted in a programming formalism that is being used in a variety of areas, but principally in those where the modelers face the issues of defaults, frame axioms and other nonmonotonic phenomena. The experience of ASP programmers shows that these phenomena can be naturally incorporated into the practice of modeling real-life problems.

We believe ASP is here to stay. It provides a venue for problem modeling, problem description and problem solving. This does not mean that the process of developing ASP is finished. Certainly new extensions of ASP will emerge in the future. Additional desiderata include: software engineering tools for testing correctness of implementation, integrated development environments and other tools that will speed up the process of the use of ASP in normal programming practice. Better grounders and better solvers able to work with incremental grounding only will certainly emerge. Similarly, new application domains will surface and bring new generations of investigators and, more importantly, users for ASP.


The work of the second author was partially supported by the Academy of Finland (project 122399). The work of the third author was partially supported by the NSF grant IIS-0913459.


  • Apt et al. (1988) K. Apt, H. A. Blair, and A. Walker. Towards a theory of declarative knowledge. In J. Minker, editor, Foundations of deductive databases and logic programming, pages 89–142. Morgan Kaufmann, 1988.
  • Balduccini et al. (2006) M. Balduccini, M. Gelfond, and M. Nogueira. Answer set based design of knowledge systems. Annals of Mathematics and Artificial Intelligence, 47(1-2):183–219, 2006.
  • Besnard (1989) P. Besnard. An Introduction to Default Logic. Springer, Berlin, 1989.
  • Bidoit and Froidevaux (1987) N. Bidoit and C. Froidevaux. Minimalism subsumes default logic and circumscription. In Proceedings of IEEE Symposium on Logic in Computer Science, LICS 1987, pages 89–97. IEEE Press, 1987.
  • Bidoit and Froidevaux (1991) N. Bidoit and C. Froidevaux. General logical databases and programs: default logic semantics and stratification. Information and Computation, 91(1):15–54, 1991.
  • Boenn et al. (2011) G. Boenn, M. Brain, M. De Vos, and J. Fitch. Automatic music composition using answer set programming. Theory and Practice of Logic Programming, 11(2-3):397-427, 2011.
  • Brain and Vos (2005) M. Brain and M. De Vos. Debugging logic programs under the answer set semantics. In Answer Set Programming, Advances in Theory and Implementation, Proceedings of the 3rd International ASP’05 Workshop, CEUR Workshop Proceedings 142, 2005.
  • Brewka (1991) G. Brewka. Nonmonotonic Reasoning: Logical Foundations of Commonsense, volume 12 of Cambridge Tracts in Theoretical Computer Science. Cambridge University Press, Cambridge, UK, 1991.
  • Brummayer and Järvisalo (2010) R. Brummayer and M. Järvisalo. Testing and debugging techniques for answer set solver development. Theory and Practice of Logic Programming, 10(4-6):741–758, 2010.
  • Cadoli and Lenzerini (1990) M. Cadoli and M. Lenzerini. The complexity of closed world reasoning and circumscription. In Proceedings of the 8th National Conference on Artificial Intelligence, pages 550–555. MIT Press, 1990.
  • Cadoli et al. (1997) M. Cadoli, T. Eiter, and G. Gottlob. Default logic as a query language. IEEE Transactions on Knowledge and Data Engineering, 9(3):448–463, 1997.
  • Chen and Warren (1996) W. Chen and D.S. Warren. Computation of stable models and its integration with logical query processing. IEEE Transactions on Knowledge and Data Engineering, 8(5):742–757, 1996.
  • Cholewiński (1995) P. Cholewiński. Stratified default theories. In L. Pacholski and J. Tiuryn, editors, Computer Science Logic, volume 933 of LNCS, pages 456–470. Springer, 1995
  • Cholewiński et al. (1996) P. Cholewiński, W. Marek, and M. Truszczyński. Default reasoning system DeReS. In Proceedings of the 5th International Conference on Principles of Knowledge Representation and Reasoning, KR 1996, pages 518–528. Morgan Kaufmann, 1996.
  • Clark (1978) K.L. Clark. Negation as failure. In H. Gallaire and J. Minker, editors, Logic and data bases, pages 293–322. Plenum Press, New York-London, 1978.
  • Dao-Tran et al. (2009) M. Dao-Tran, T. Eiter, M. Fink, and T. Krennwallner. Modular nonmonotonic logic programming revisited. In P. M. Hill and D. S. Warren, editors, Proceedings of the 25th International Conference on Logic Programming, ICLP 2009, volume 5649 of LNCS, pages 145–159. Springer, 2009.
  • Denecker (1998) M. Denecker. The well-founded semantics is the principle of inductive definition. In J. Dix, L. Fariñas del Cerro, and U. Furbach, editors, Proceedings of the 6th European Workshop on Logics in Artificial Intelligence, JELIA 1998, volume 1489 of LNCS, pages 1–16. Springer, 1998.
  • Denecker (2000) M. Denecker. Extending classical logic with inductive definitions. In J. W. Lloyd, V. Dahl, U. Furbach, M. Kerber, K.-K. Lau, C. Palamidessi, L. Moniz Pereira, Y. Sagiv, P. J. Stuckey, editors, Computational Logic, volume 1861 of LNCS, pages 703–717. Springer, 2000.
  • Denecker and Ternovska (2008) M. Denecker and E. Ternovska. A logic for non-monotone inductive definitions. ACM Transactions on Computational Logic, 9(2), 2008.
  • Denecker et al. (2003) M. Denecker, V.W. Marek, and M. Truszczyński. Uniform semantic treatment of default and autoepistemic logics. Artificial Intelligence Journal, 143:79–122, 2003.
  • Denecker et al. (2009) M. Denecker, J. Vennekens, S. Bond, M. Gebser, and M. Truszczynski. The second answer set programming competition. In E. Erdem, F. Lin, and T. Schaub, editors, Proceedings of the 10th International Conference on Logic Programming and Nonmonotonic Reasoning, LPNMR 2009, volume 5753 of LNCS, pages 637–654. Springer, 2009.
  • Denecker et al. (2011) M. Denecker, V.W. Marek, and M. Truszczynski. Reiter’s Default Logic Is a Logic of Autoepistemic Reasoning And a Good One, Too. This volume.
  • Dowling and Gallier (1984) W.F. Dowling and J.H. Gallier. Linear-time algorithms for testing the satisfiability of propositional Horn formulae. Journal of Logic Programming, 1(3):267–284, 1984.
  • East and Truszczyński (2006) D. East and M. Truszczyński. Predicate-calculus based logics for modeling and solving search problems. ACM Transactions on Computational Logic, 7:38–83, 2006.
  • Eiter and Gottlob (1993a) T. Eiter and G. Gottlob. Propositional circumscription and extended closed world reasoning are -complete. Theoretical Computer Science, 114(2):231 – 245, 1993a.
  • Eiter and Gottlob (1993b) T. Eiter and G. Gottlob. Complexity results for disjunctive logic programming and application to nonmonotonic logics. In D. Miller, editor, Proceedings of the 1993 International Symposium on Logic Programming, pages 266–278. MIT Press, 1993b.
  • Eiter and Gottlob (1995) T. Eiter and G. Gottlob. On the computational cost of disjunctive logic programming: propositional case. Annals of Mathematics and Artificial Intelligence, 15(3-4):289–323, 1995.
  • Erdem (2011) E. Erdem. Applications of answer set programming in phylogenetic systematics. In M. Balduccini and T.C. Son, editors, Essays in Honor of Michael Gelfond. Springer, 2011.
  • Erdem and Lifschitz (2003) E. Erdem and V. Lifschitz. Tight logic programs. Theory and Practice of Logic Programming, 3(4-5):499–518, 2003.
  • Etherington (1987) D. W. Etherington. Formalizing nonmonotonic reasoning systems. Artificial Intelligence, 31(1):41–85, 1987.
  • Finger (1987) J. J. Finger. Exploiting Constraints in Design Synthesis. PhD thesis, Stanford University, 1987.
  • Finkel and O’Sullivan (2011) R. A. Finkel and B. O’Sullivan. Reasoning about conditional constraint specifications. Artificial Intelligence for Engineering Design, Analysis and Manufacturing, 25(2):163-174, 2011.
  • Fitting (1985) M. C. Fitting. A Kripke-Kleene semantics for logic programs. Journal of Logic Programming, 2(4):295–312, 1985.
  • Friedrich et al. (2010) G. Friedrich, M. Fugini, E. Mussi, B. Pernici, and G. Tagni. Exception handling for repair in service-based processes. IEEE Transactions on Software Engineering, 36(2):198–215, 2010.
  • Gabbay (1989) D.M. Gabbay. Theoretical foundations for non-monotonic reasoning in expert systems. In Proceedings of the NATO Advanced Study Institute on Logics and Models of Concurrent Systems, pages 439–457. Springer, 1989.
  • Gebser et al. (2007) M. Gebser, L. Liu, G. Namasivayam, A. Neumann, T. Schaub, and M. Truszczyński. The first answer set programming system competition. In C. Baral, G. Brewka, and J. Schlipf, editors, Proceedings of the 9th International Conference on Logic Programming and Nonmonotonic Reasoning, LPNMR 2007, volume 4483 of LNCS, pages 3–17. Springer, 2007.
  • Gebser et al. (2010a) M. Gebser, C. Guziolowski, M. Ivanchev, T. Schaub, A. Siegel, S. Thiele, and P. Veber. Repair and prediction (under inconsistency) in large biological networks with answer set programming. In F. Lin, U. Sattler and M. Truszczynski, editors, Proceedings of the 12th International Conference on Principles of Knowledge Representation and Reasoning, KR 2010, pages 497–507. AAAI Press, 2010a.
  • Gebser et al. (2010b) M. Gebser, A. König, T. Schaub, S. Thiele, and P. Veber. The BioASP library: ASP solutions for systems biology. In Proceedings of the 22nd IEEE International Conference on Tools with Artificial Intelligence, ICTAI 2010, pages 383–389, 2010b.
  • Gelfond (1987) M. Gelfond. On stratified autoepistemic theories. In Proceedings of AAAI 1987, pages 207–211. Morgan Kaufmann, 1987.
  • Gelfond and Lifschitz (1988) M. Gelfond and V. Lifschitz. The stable model semantics for logic programming. In R. A. Kowalski and K. A. Bowen, editors, Proceedings of the International Joint Conference and Symposium on Logic Programming, pages 1070–1080. MIT Press, 1988.
  • Gelfond and Lifschitz (1991) M. Gelfond and V. Lifschitz. Classical negation in logic programs and disjunctive databases. New Generation Computing, 9:365–385, 1991.
  • Ginsberg (1989) M.L. Ginsberg. A circumscriptive theorem prover. In M. Reinfrank, J. de Kleer, M. L. Ginsberg, and E. Sandewall, editors, Proceedings of the 2nd International Workshop on Non-Monotonic Reasoning, volume 346 of LNCS, pages 100–114. Springer, 1989.
  • Gottlob (1992) G. Gottlob. Complexity results for nonmonotonic logics. Journal of Logic and Computation, 2(3):397–425, 1992.
  • Hanks and McDermott (1986) S. Hanks and D. McDermott. Default reasoning, nonmonotonic logics and frame problem. In Proceedings of AAAI 1986, pages 328–333. Morgan Kaufmann, 1986.
  • Janhunen et al. (2009) T. Janhunen, E. Oikarinen, H. Tompits, and S. Woltran. Modularity aspects of disjunctive stable models. Journal of Artificial Intelligence Research, 35:813–857, 2009.
  • Kautz and Selman (1989) H.A. Kautz and B. Selman. Hard problems for simple default logics. In R. J. Brachman, H. J. Levesque, and R. Reiter, editors, Proceedings of the 1st International Conference on Principles of Knowledge Representation and Reasoning, KR 1989, pages 189–197. Morgan Kaufmann, 1989.
  • Konolige (1988) K. Konolige. On the relation between default and autoepistemic logic. Artificial Intelligence, 35(3):343–382, 1988.
  • Konolige (1989) K. Konolige. Errata: On the relation between default and autoepistemic logic. Artificial Intelligence, 41(1):115, 1989.
  • Kraus et al. (1990) S. Kraus, D. Lehmann, and M. Magidor. Nonmonotonic reasoning, preferential models and cumulative logics. Artificial Intelligence, 44:167–207, 1990.
  • Kunen (1987) K. Kunen. Negation in logic programming. Journal of Logic Programming, 4(4):289–308, 1987.
  • Lehmann and Magidor (1992) D. Lehmann and M. Magidor. What does a conditional knowledge base entail? Artificial Intelligence, 55:1–60, 1992.
  • Leone et al. (2006) N. Leone, G. Pfeifer, W. Faber, T. Eiter, G. Gottlob, S. Perri, and F. Scarcello. The DLV system for knowledge representation and reasoning. ACM Transactions on Computational Logic, 7(3):499–562, 2006.
  • Lierler and Maratea (2004) Y. Lierler and M. Maratea. Cmodels-2: SAT-based answer set solver enhanced to non-tight programs. In V. Lifschitz, I. Niemelä, editors, Proceedings of the 7th International Conference on Logic Programming and Nonmonotonic Reasoning, LPNMR 2004, volume 2923 of LNCS, pages 346–350. Springer, 2004.
  • Lifschitz (1988) V. Lifschitz. Circumscriptive theories: a logic-based framework for knowledge representation. Journal of Philosophical Logic, 17(4):391–441, 1988.
  • Lifschitz and Turner (1994) V. Lifschitz and H. Turner. Splitting a logic program. In P. Van Hentenryck, editor, Proceedings of the 11th International Conference on Logic Programming, ICLP 1994, pages 23–37. MIT Press, 1994.
  • Lin and Zhao (2002) F. Lin and Y. Zhao. ASSAT: Computing answer sets of a logic program by SAT solvers. In Proceedings of the 18th National Conference on Artificial Intelligence, AAAI 2002, pages 112–117. AAAI Press, 2002.
  • Liu and Truszczyński (2005) L. Liu and M. Truszczyński. Pbmodels - software to compute stable models by pseudoboolean solvers. In C. Baral, G. Greco, N. Leone, G. Terracina Proceedings of the 8th International Conference on Logic Programming and Nonmonotonic Reasoning, LPNMR 2005, volume 3662 of LCNS, pages 410–415. Springer, 2005.
  • Makinson (1989) D. Makinson. General theory of cumulative inference. In M. Reinfrank, J. de Kleer, M. L. Ginsberg, and E. Sandewall, editors, Proceedings of the 2nd International Workshop on Non-Monotonic Reasoning, volume 346 of LNCS, pages 1–18. Springer, 1989.
  • Marek and Remmel (2003) V.W. Marek and J.B. Remmel. On the expressibility of stable logic programming. Theory and Practice of Logic Programming, 3:551–567, 2003.
  • Marek and Truszczyński (1999) V.W. Marek and M. Truszczyński. Stable models and an alternative logic programming paradigm. In K.R. Apt, W. Marek, M. Truszczyński, and D.S. Warren, editors, The Logic Programming Paradigm: a 25-Year Perspective, pages 375–398. Springer, 1999.
  • Marek (1989) W. Marek. Stable theories in autoepistemic logic. Fundamenta Informaticae, 12(2):243–254, 1989.
  • Marek and Truszczyński (1989) W. Marek and M. Truszczyński. Relating autoepistemic and default logics. In R. J. Brachman, H. J. Levesque, and R. Reiter, editors, Proceedings of the 1st International Conference on Principles of Knowledge Representation and Reasoning, KR 1989, pages 276–288. Morgan Kaufmann, 1989.
  • Marek and Truszczyński (1989b) W. Marek and M. Truszczyński. Stable semantics for logic programs and default theories. In E.Lusk and R. Overbeek, editors, Proceedings of the North American Conference on Logic Programming, pages 243–256. MIT Press, 1989b.
  • Marek and Truszczyński (1991) W. Marek and M. Truszczyński. Autoepistemic logic. Journal of the ACM, 38(3):588–619, 1991.
  • Marek and Truszczyński (1993) W. Marek and M. Truszczyński. Nonmonotonic Logic; Context-Dependent Reasoning. Springer, 1993.
  • Marek et al. (1993) W. Marek, G.F. Shvarts, and M. Truszczyński. Modal nonmonotonic logics: ranges, characterization, computation. Journal of the ACM, 40(4):963–990, 1993.
  • Marek et al. (1994) W. Marek, A. Nerode, and J. B. Remmel. The stable models of predicate logic programs. Journal of Logic Programming, 21(3):129–154, 1994.
  • McCarthy (1977) J. McCarthy. Epistemological problems of Artificial Intelligence. In Proceedings of the 5th International Joint Conference on Artificial Intelligence, pages 1038–1044, 1977.
  • McCarthy (1980) J. McCarthy. Circumscription — a form of non-monotonic reasoning. Artificial Intelligence, 13(1-2):27–39, 1980.
  • McCarthy and Hayes (1969) J. McCarthy and P. Hayes. Some philosophical problems from the standpoint of artificial intelligence. In B. Meltzer and D. Michie, editors, Machine Intelligence 4, pages 463–502. Edinburgh University Press, 1969.
  • McDermott (1982) D. McDermott. Nonmonotonic logic II: nonmonotonic modal theories. Journal of the ACM, 29(1):33–57, 1982.
  • McDermott and Doyle (1980) D. McDermott and J. Doyle. Nonmonotonic logic I. Artificial Intelligence, 13(1-2):41–72, 1980.
  • Moore (1984) R. C. Moore. Possible-world semantics for autoepistemic logic. In R. Reiter, editor, Proceedings of the Workshop on Non-Monotonic Reasoning, pages 344–354, 1984. Reprinted in: M. Ginsberg, editor, Readings on Nonmonotonic Reasoning, pages 137–142, Morgan Kaufmann, 1990.
  • Moore (1985) R.C. Moore. Semantical considerations on nonmonotonic logic. Artificial Intelligence, 25(1):75–94, 1985.
  • Niemelä (1988) I. Niemelä. On the complexity of the decision problem in propositional nonmonotonic logic. In E. Börger, H. Kleine Büning, M. M. Richter, editors, Proceedings of the 2nd Workshop on Computer Science Logic, CSL 1988, volume 385 of LNCS, pages 226–239. Springer, 1988.
  • Niemelä (1988) I. Niemelä. Decision procedure for autoepistemic logic. In R. Overbeek E. Lusk, editor, Proceedings of the 9th International Conference on Automated Deduction, volume 310 of LNCS, pages 675–684. Springer, 1988.
  • Niemelä (1990) I. Niemelä. Towards automatic autoepistemic reasoning. In J. van Eijck, editor, Proceedings of the European Workshop on Logics in Artificial Intelligence, JELIA 1990, volume 478 of LNCS, pages 428–443. Springer, 1990.
  • Niemelä (1992) I. Niemelä. On the decidability and complexity of autoepistemic reasoning. Fundamenta Informaticae, 17(1-2):117–155, 1992.
  • Niemelä (1994) I. Niemelä. A decision method for nonmonotonic reasoning based on autoepistemic reasoning. In J. Doyle, E. Sandewall and P. Torasso, Proceedings of the 4th International Conference on Principles of Knowledge Representation and Reasoning, pages 473–484. Morgan Kaufmann, 1994.
  • Niemelä (1995) I. Niemelä. Towards efficient default reasoning. In C. S. Mellish, editor, Proceedings of IJCAI 1995, pages 312–318. Morgan Kaufmann, 1995.
  • Niemelä (1998) I. Niemelä. Logic programs with stable model semantics as a constraint programming paradigm. In Proceedings of the Workshop on Computational Aspects of Nonmonotonic Reasoning, pages 72–79. Helsinki University of Technology, Digital Systems Laboratory, Research Report A52, May 1998.
  • Niemelä (1999) I. Niemelä. Logic programming with stable model semantics as a constraint programming paradigm. Annals of Mathematics and Artificial Intelligence, 25(3-4):241–273, 1999.
  • Niemelä and Rintanen (1992) I. Niemelä and J. Rintanen. On the impact of stratification on the complexity of nonmonotonic reasoning. In B. Nebel, C. Rich, and W. R. Swartout, editors, Proceedings of the 3rd International Conference on Principles of Knowledge Representation and Reasoning, pages 627–638. Morgan Kaufmann, 1992.
  • Niemelä and Simons (1995) I. Niemelä and P. Simons. Evaluating an algorithm for default reasoning. In Working Notes of the IJCAI’95 Workshop on Applications and Implementations of Nonmonotonic Reasoning Systems, Montreal, Canada, pages 66–72, Montreal, Canada, August 1995.
  • Niemelä and Simons (1996) I. Niemelä and P. Simons. Efficient implementation of the well-founded and stable model semantics. In M. Maher, editor, Proceedings of the Joint International Conference and Symposium on Logic Programming, pages 289–303. MIT Press, 1996.
  • Niemelä and Tuominen (1986) I. Niemelä and H. Tuominen. A system for logical expertise. In Proceedings of the Finnish Artificial Intelligence Symposium, Volume 2, pages 44–53, Espoo, Finland, August 1986. Finnish Society of Information Processing Science, 1986.
  • Niemelä and Tuominen (1987) I. Niemelä and H. Tuominen. Helsinki Logic Machine: a system for logical expertise. Technical report B1, Helsinki University of Technology, Digital Systems Laboratory, Espoo, Finland, December 1987.
  • Niemelä et al. (1999) I. Niemelä, P. Simons, and T. Soininen. Stable model semantics of weight constraint rules. In M. Gelfond, N. Leone, and G. Pfeifer, editors, Proceedings of the 5th International Conference on Logic Programming and Nonmonotonic Reasoning, LPNMR 1999, volume 1730 of LNCS, pages 317–331. Springer, 1999.
  • Pearl (1990) J. Pearl. System Z: A natural ordering of defaults with tractable applications to nonmonotonic reasoning. In R. Parikh, editor, Proceedings of the 3rd Conference on Theoretical Aspects of Reasoning about Knowledge, TARK 1990, pages 121–135. Morgan Kaufmann, 1990.
  • Reiter (1978) R. Reiter. On closed world data bases. In H. Gallaire and J. Minker, editors, Logic and data bases, pages 55–76. Plenum Press, 1978.
  • Reiter (1980) R. Reiter. A logic for default reasoning. Artificial Intelligence, 13(1-2):81–132, 1980.
  • Reiter and Criscuolo (1981) R. Reiter and G. Criscuolo. On interacting defaults. In P. J. Hayes, editor, Proceedings of IJCAI 1981, pages 270–276. William Kaufman, 1981.
  • Schlipf (1995) J. Schlipf. The expressive powers of the logic programming semantics. Journal of Computer and System Sciences, 51(1):64–86, 1995.
  • Schwarz (1991) G.F. Schwarz. Autoepistemic logic of knowledge. In A. Nerode, W. Marek, and V.S. Subrahmanian, editors, Proceedings of the 1st International Workshop on Logic Programming and Nonmonotonic Reasoning, LPNMR 1991, pages 260–274. MIT Press, 1991.
  • Schwarz (1992) G.F. Schwarz. Minimal model semantics for nonmonotonic modal logics. In Proceedings of LICS 1992, pages 34–43. IEEE Computer Society, 1992.
  • Schwarz and Truszczyński (1992) G.F. Schwarz and M. Truszczyński. Modal logic S4F and the minimal knowledge paradigm. In Y. Moses, editor, Proceedings of TARK 1992, pages 34–43. Morgan Kaufmann, 1992.
  • Shvarts (1990) G.F. Shvarts. Autoepistemic modal logics. In R. Parikh, editor, Proceedings of TARK 1990, pages 97–109, Morgan Kaufmann, 1990.
  • Simons (1999) P. Simons. Extending the stable model semantics with more expressive rules. In M. Gelfond, N. Leone, and G. Pfeifer, editors, Proceedings of the 5th International Conference on Logic Programming and Nonmonotonic Reasoning, LPNMR 1999, volume 1730 of LNCS, pages 305–316. Springer, 1999.
  • Soininen and Niemelä (1998) T. Soininen and I. Niemelä. Developing a declarative rule language for applications in product configuration. In G. Gupta, editor, Proceedings of the 1st International Workshop on Practical Aspects of Declarative Languages, PADL 1999, volume 1551 of LNCS, pages 305–319. Springer, 1998.
  • Stalnaker (1980) R.C. Stalnaker. A note on nonmonotonic modal logic. Artificial Intelligence, 64(2):183–196, 1993; broadly available as unpublished manuscript, 1980.
  • Stillman (1992) J. Stillman. The complexity of propositional default logics. In W. R. Swartout, editor, Proceedings of the 10th National Conference on Artificial Intelligence, AAAI 1992, pages 794–800, MIT Press, 1992.
  • Tiihonen et al. (2003) J. Tiihonen, T. Soininen, I. Niemelä, and R. Sulonen. A practical tool for mass-customizing configurable products. In Proceedings of the 14th International Conference on Engineering Design, pages 1290–1299, 2003.
  • Truszczyński (1991) M. Truszczyński. Modal interpretations of default logic. In R. Reiter and J. Mylopoulos, editors, Proceedings of IJCAI 1991, pages 393–398. Morgan Kaufmann, 1991.
  • Van Gelder et al. (1991) A. Van Gelder, K.A. Ross, and J.S. Schlipf. The well-founded semantics for general logic programs. Journal of the ACM, 38(3):620–650, 1991.