1 Introduction
Programming autonomous agents that behave intelligently is one of the key challenges of Artificial Intelligence. Because of its declarative nature, and high level of abstraction, logic programming is a natural choice for this task. This is witnessed by e.g. the two major exponents of agent programming languages that are based on classical logic programming, namely Golog
[Levesque et al. (1997)] and Flux [Thielscher (2005a)].Both these languages combine a language for specifying the agent’s behaviour with an axiomatic theory that describes the agent’s environment. In the case of Golog the strategy language is procedural in nature (though implemented in Prolog), and the action theory is the classical Situation Calculus [McCarthy and Hayes (1969)] in Reiter’s version [Reiter (2001a)]. For Flux the strategy language is full classical logic programming, and the action theory is the more recent Fluent Calculus [Thielscher (1999)].
In a recent work [Drescher et al. (2009)] we have developed Agent Logic Programs (ALPs), a new declarative strategy language that is based upon a proof calculus in the style of classical SLDresolution. Contrary to Golog and Flux the ALP framework is parametric in the action theory: any background theory that allows to infer when an action is applicable, and what the effects of the action are, can be used. Exploiting this generality we have recently [Thielscher (2010b)] been able to give a semantics for the BDIstyle language AgentSpeak [Bordini et al. (2007)]. Another distinctive feature of the theoretical framework is the elegant handling of incomplete information for offline planning via disjunctive substitutions. By default, ALPs are combined with our new Unifying Action Calculus (UAC) [Thielscher (2011)] that encompasses the major logical action calculi, including both the Situation Calculus and the Fluent Calculus, as well as many planning domain description languages. The ALP formalism stays entirely within classical logic.
The implementation of any fragment of the ALPprolog framework consists of (1) an implementation of the proof calculus, and (2) an action theory reasoner. Existing mature Prolog technology can be used out of the box for (1) unless disjunctive substitutions enter the picture. For (2) we can also exploit existing technology: E.g. Golog implements a fragment of the Situation Calculus, and Flux handles a fragment of the Fluent Calculus. In [Drescher et al. (2009)] the implementation of a Description Logicbased fragment of the Fluent Calculus is described.
In this work we present ALPprolog, where the underlying action theory is an essentially propositional version of the Fluent Calculus in the UAC that includes a simple, yet powerful model of sensing. ALPprolog is intended for the online control of agents, where actions are immediately executed. This starkly contrasts with offline reasoning, where agents may make assumptions to see where these are leading. ALPprolog was developed specifically for the efficient handling of large ground state representations, something that we consider to be practically useful. To this end ALPprolog combines strongpoints of Golog and Flux:

From Golog it takes the representation of the agent’s state knowledge in full propositional logic via prime implicates; and

From Flux it takes the principle of progression: The agent’s state knowledge is updated upon the execution of an action. In standard Golog the agent’s initial state knowledge is never updated.^{1}^{1}1But there is a version of Golog where the initial state is periodically updated [Sardina and Vassos (2005)]. Instead, queries referring to later timepoints are rewritten until they can be evaluated against the initial state knowledge, something which becomes a hindrance to good performance as the sequence of executed actions grows.
We emphasise that ALPprolog is an agent programming language in the spirit of classical logic programming in Prolog: The straightforward operational semantics provides the programmer with a powerful means to actively determine the sequence of actions that an agent executes. ALPprolog^{2}^{2}2The name is a play on ALPs, propositional logic, and the implementation in plain Prolog. can be obtained at alpprolog.sourceforge.net.
2 ALPs in a Nutshell
The purpose of agent logic programs is to provide highlevel control
programs for agents using a combination of declarative programming
with reasoning about actions. The syntax of these programs is
kept very simple: standard (definite) logic programs (see e.g. [J.W. Lloyd (1987)]) are augmented
with just two
special predicates, one — written do(
)
— to
denote the execution of an action by the agent, and one — written
?(
)
— to verify properties against (the
agent’s model of)
the state of its environment. This model, and how it is affected by
actions, is defined in a separate action theory. This allows for a
clear separation between the agent’s strategic behaviour (given by the
agent logic program itself) and the underlying theory about the
agent’s actions and their effects. Prior to giving the formal
definition, let us illustrate the idea by an example agent
logic program.
Example 1
Consider an agent whose task is to find gold in a maze. For the sake of simplicity, the states of the environment shall be described by a single fluent (i.e., state property): to denote that is at location . The agent can perform the action of going to location , which is possible if is adjacent to the current location of the agent. The following ALP describes a simple search strategy via a given list of locations (choice points) that the agent may visit, and an ordered collection of backtracking points. We follow the Prolog convention of writing variables with a leading uppercase letter.
explore(Choicepoints,Backtrack) : % finished, if ?(at(agent,X)), ?(at(gold,X)). % gold is found explore(Choicepoints,Backtrack) : ?(at(agent,X)), select(Y,Choicepoints,NewChoicepoints), % choose a direction do(go(Y)), % go in this direction explore(NewChoicepoints,[XBacktrack]). % store the choice explore(Choicepoints,[XBacktrack]) : % go back one step do(go(X)), explore(Choicepoints,Backtrack). select(X,[XXs],Xs). select(X,[YXs],[YYs]) : select(X,Xs,Ys).
Suppose we are given
a
list of choice points C
,
then the query : explore(C,[])
lets the agent
systematically search for gold from its current location: the first
clause describes the base case where the agent is successful; the
second clause lets the agent select a new location from the list of
choice points and go to this location (the declarative semantics and
proof theory for do(
)
will require that the
action is possible at the time of execution); and the
third clause sends the agent back using the latest backtracking
point.
The example illustrates two distinct features of ALPs: (1) The agent strategy is defined by a logic program that may use arbitrary function and predicate symbols in addition to the signature of the underlying action theory. (2) The update of the agent’s belief according to the effects of its actions is not part of the strategy. Formally, ALPs are defined as follows.
Definition 1
Consider an action theory signature that includes the predefined sorts action and fluent, along with a logic program signature .

Terms are from .

If
p
is an ary relation symbol from and tt are terms, then p(tt is a program atom. 
do(
)
is a program atom if is an action term in . 
?(
)
is a program atom if is a state property in , that is, a formula (represented as a term) based on the fluents in . 
Clauses, programs, and queries are then defined as usual for definite logic programs, with the restriction that the two special atoms cannot occur in the head of a clause.
2.1 Declarative Semantics: Program + Action Theory
The semantics of an ALP is given in two steps. First,
the program needs to be “temporalised,” making explicit
the state change that is implicit in the
use of the two special predicates,
do(
)
and
?(
)
.
Second, the resulting program is combined with an action
theory as the basis for evaluating these two special predicates.
The semantics is then the classical logical semantics of the expanded program together with the action theory.
Time is incorporated into a program through macroexpansion:
two arguments of sort time^{3}^{3}3Which specific concept of
time is being
used depends on how the sort time is defined in the
underlying action theory, which may be
branching (as, e.g., in the Situation Calculus) or
linear (as, e.g., in the Event Calculus).
are added to every regular
program atom , and then is understood as
restricting the truth of the atom to the temporal interval between
(and including) and .
The two special atoms receive special
treatment: ?(
)
is rewritten to
, with the intended meaning that
is true at ; and do(
)
is mapped onto , meaning that action can
be executed at and that its execution ends in .
The formal definition is as follows.
Definition 2
For a clause H : B,...,B (), let be variables of sort time.

For , if B is of the form

p(t,...,t), expand to .

do(), expand to .

?(), expand to .


The head atom H = p(t,...,t) is expanded to .

The resulting clauses are understood as universally quantified implications.
Queries are expanded exactly like clause bodies, except that

a special constant — denoting the earliest timepoint in the underlying action theory — takes the place of ;

the resulting conjunction is existentially quantified.
Example 1
The example program of the preceding section is understood as the following axioms, which for notational convenience we have simplified in that all equations between time variables have been applied and then omitted.
The resulting theory constitutes a purely logical axiomatisation of the
agent’s strategy, which provides the basis for logical entailment.
For instance, macroexpanding the
query : explore(C,[])
from the above example results in the temporalised logical formula
.
If this formula follows from
the axioms above, then that means that the strategy can be
successfully executed, starting at , for the given list of
choice points . Whether this is actually the case of course
depends on the additional action theory that is needed to evaluate
the special atoms Holds and Poss in a macroexpanded program.
Macroexpansion provides the first part of the declarative semantics of an agent logic program; the second part is given by an action theory in form of a logical axiomatisation of actions and their effects. The overall declarative semantics of agent logic programs is given by the axiomatisation consisting of the action theory and the expanded program.
Let us next introduce the fragment of the UAC corresponding to the Fluent Calculus. The UAC that is used to axiomatise the action theory is based on manysorted first order logic with equality and the four sorts time, fluent, object, and action. By convention variable symbols , , , and are used for terms of sort time, fluent, object, and action, respectively. Fluents are reified, and the standard predicate indicates whether a fluent is true at a particular time. The predicate means that action can be executed at and that its execution ends in . The number of function symbols into sorts fluent and action is finite.
Definition 3 (Action Theory Formula Types)
We stipulate that the following formula types are used by action theories:

State formulas express what is true at particular times: A state formula in is a firstorder formula with free variables where

for each occurrence of we have ;

predicate Poss does not occur.
A state formula is pure if it does not mention predicates other than Holds.


A state property is an expression built from the standard logical connectives and terms of sort fluent. With a slight abuse of notation, by we denote the state formula obtained from state property by replacing every occurrence of a fluent by . In an expanded program we always treat as atomic. State properties are used by agent logic programs in
?(Phi)
atoms. 
The initial state axiom is a state formula in , where denotes the initial situation.

An action precondition axiom is of the form
where is a state formula in with free variables among . This axiom illustrates how different actions lead to different situation terms . Situations constitute the sort time in the Fluent Calculus and provide a branching time structure.

Effect axioms are of the form
Such an effect axiom has different cases that can apply — these are identified by the case selection formulas which are state formulas in with free variables among . The (and , respectively) are fluent terms with variables among and describe the positive (or, respectively, negative) effects of the action, given that case applies.

Domain constraints are universally quantified state formulas in .

Auxiliary axioms are domaindependent, but timeindependent, additional axioms such as e.g. an axiomatisation of finite domain constraints.
An action theory is given by an initial state axiom , finite sets and of precondition and effect axioms. Moreover domain constraints and auxiliary axioms may be included. For illustration, the following is a background axiomatisation for our example scenario as a basic Fluent Calculus theory in the UAC.
Example 1
Our example program can be supported by the following domain theory.

Initial state axiom

Precondition axiom

Effect axiom
Given this (admittedly very simple, for the sake of illustration) specification of the background action theory, the axiomatisation of the agent’s strategy from above entails, for example, . This can be shown as follows. First, observe that the background theory entails
where denotes the situation term . It follows that according to the first clause of our example ALP. Consider, now, the situation , then action theory and strategy together imply
By using this in turn, along with from above, according to the second program clause we obtain . Continuing this line of reasoning, it can be shown that
This proves the claim that . On the other hand e.g. the query is not entailed under the given background theory: Without location among the choice points, the strategy does not allow the agent to reach the only location that is known to house gold.
2.2 Operational Semantics: Proof Calculi
We have developed two sound and complete proof calculi for ALPs that both assume the existence of a suitable reasoner for the underlying action theory [Drescher et al. (2009)].
The first proof calculus is plain SLDresolution, only that Holds and Possatoms are evaluated against the action theory. This calculus is sound and complete if the underlying action theory has the witness property: That is, whenever then there is a substitution such that . Note that in general action theories may violate the witness property, as they may include disjunctive or purely existential information; consider e.g. the case , where the exact location of the gold is unknown.
Hence the second proof calculus, intended for the general case, resorts to constraint logic programming, and the notion of a disjunctive substitution: Still assuming that the gold is located at one of two locations the query can now be answered positively via the disjunctive substitution . Disjunctive substitution together with the respective principle of reasoning by cases are a powerful means for inferring conditional plans.
For the online control of agents, however, assuming a particular case is unsafe. But if we use the plain SLDresolutionbased ALP proof calculus on top of action theories that lack the witness property we obtain a nice characterisation of cautious behaviour in a world of unknowns (albeit at the cost of sacrificing logical completeness). For ALPprolog this is the setting that we use.
In both proof calculi we adopt the ”leftmost” computation rule familiar from Prolog. This has many advantages: First, it simplifies the implementation, as this can be based on existing mature Prolog technology. Second, state properties can always be evaluated against a description of the ”current” state. Last, but not least, this ensures that actions are executed in the order intended by the programmer — this is of no small importance for the online control of agents.
3 ALPprolog
We next present ALPprolog — an implementation of the ALP framework atop of action theories in a version of the Fluent Calculus that

uses (a notational variant of) propositional logic for describing state properties;

is restricted to actions with ground deterministic effects; and

includes sensing actions.
The intended application domain for ALPprolog is the online control of agents in dynamic domains with incomplete information.
3.1 ALPprolog Programs
An ALPprolog program is an ALP that respects the following restrictions on the ?(Phi)
atoms in the program:

All occurrences of nonfluent expressions in are positive.

So called sense fluents that represent the interface to a sensor may only occur in the form
?(s(X))
. Sense fluents are formally introduced below.
Because ALPprolog programs are meant for online execution the programmer must ensure that no backtracking over action executions occurs, by inserting cuts after all action occurrences. Observe that this applies to sensing actions, too. It is readily checked that — after the insertion of cuts — the ALP from example 1 satisfies all of the above conditions.
3.2 Propositional Fluent Calculus
In this section we introduce the announced propositional fragment of the Fluent Calculus. The discussion of sensing is deferred until section 3.3.
For ease of modelling we admit finitely many ground terms for fluents and objects, instead of working directly with propositional letters. An action domain is then made propositional by including the respective domain closure axioms. For actions, objects, and fluents unique name axioms are included — hence we can avoid equality reasoning.
The basic building block of both the propositional Fluent Calculus and ALPprolog are the socalled prime implicates of a state formula :
Definition 4 (Prime Implicate)
A clause is a prime implicate of iff it is entailed by , is not a tautology, and is not entailed by another prime implicate.
The prime implicates of a formula are free from redundancy — all tautologies and implied clauses have been deleted. For any state formula an equivalent prime state formula can be obtained by first transforming the state formula into a set of clauses, and by then closing this set under resolution, and the deletion of subsumed clauses and tautologies.
Prime state formulas have the following nice property: Let be a prime state formula, and let be some clause (not mentioning auxiliary predicates); then is entailed by if and only if it is subsumed by some prime implicate in , a fact that has already been exploited for Golog [Reiter (2001a), Reiter (2001b)]. This property will allow us to reduce reasoning about state knowledge in ALPprolog to simple list lookup operations.
Formally the propositional version of the Fluent Calculus is defined as follows.
Definition 5 (Propositional Fluent Calculus Domain)
We stipulate that the following properties hold in propositional Fluent Calculus domains:

The initial state is specified by a ground prime state formula.

The state formulas in action preconditions are prime state formulas.

The effect axioms are of the form
where each is a prime state formula. This implies that existentially quantified variables that may occur in case selection formulas (cf. definition 3) have been eliminated by introducing additional cases.

Only socalled modular domain constraints [Herzig and Varzinczak (2007)] may be included. Very roughly, domain constraints are modular if they can be compiled into the agent’s initial state knowledge, and the effect axioms ensure that updated states also respect the domain constraints. In the Fluent Calculus this holds if the following two conditions are met [Thielscher (2011)]: Condition (1), says that for a state that is consistent with the domain constraints and in which an action is applicable, the condition for at least one case in the effect axiom for holds. Condition (2) requires that any possible update leads to a state that satisfies the domain constraints. Formally, let be constants of sort Time. the domain constraints, the precondition axioms, and the effect axioms. The following must hold for every action : There exists such that
(1) and for every such ,
(2) Nonmodular, fully general domain constraints greatly complicate reasoning.

Auxiliary timeindependent axioms may be included if they can faithfully be represented in the Prolog dialect underlying the implementation. This deliberately sloppy condition is intended to allow the programmer to use her favourite Prolog library. However, we stipulate that auxiliary predicates occur only positively outside of in the action domain in order to ensure that they can safely be evaluated by Prolog. They also must not occur in the initial state formula at all. The update mechanism underlying ALPprolog can handle only ground effects. Hence, if auxiliary atoms are used in action preconditions, case selection formulas of effect axioms, then it is the burden of the programmer to ensure that these predicates always evaluate to ground terms on those variables that also occur in the action’s effects.
On the one hand clearly every propositional Fluent Calculus domain can be transformed to this form. On the other hand it is well known that in general compiling away the quantifiers in a state formula can result in an exponential blowup, as can the conversion to conjunctive normal form. We believe that the simplicity of reasoning with prime implicates outweighs this drawback.
Propositional action domains can still be nondeterministic. For example, for an applicable action two different cases may be applicable at the same time. The resulting state would then be determined only by the disjunction of the cases’ effects. What is more, it would be logically unsound to consider only the effects of one of the cases. For the online control of agents in ALPprolog we stipulate that for an applicable action at most a single case applies, greatly simplifying the update of the agent’s state knowledge.
Definition 6 (Deterministic Propositional Fluent Calculus)
A propositional Fluent Calculus domain is deterministic if the following holds: Let be an applicable ground action. Then there is at most one case of the action that is applicable in the given state.
For example, an action theory is deterministic if for each effect axiom all the cases are mutually exclusive. Next assume we have an applicable deterministic action with e.g. two case selection formulas and , where neither case is implied by the current state. Here, instead of updating the current state with the disjunction of the respective effects, ALPprolog will employ incomplete reasoning.
3.3 Propositional Fluent Calculus with Sensing
We make the following assumptions concerning sensing: At any one time, a sensor may only return a single value from a fixed set of ground terms, the sensing results. However, the meaning of such a sensing result may depend upon the concrete situation of the agent.
Example 1
Assume that now one of the cells in the maze contains a deadly threat to our goldhunting agent. If the agent is next to a cell containing the threat she perceives a certain smell, otherwise she doesn’t: She can sense whether one of the neighbouring cells is unsafe; but the actual neighbouring cells are only determined by the agent’s current location.
Definition 7 (Sensor Axiom)
A sense fluent is a unary fluent that serves as interface to the sensor. We assume the sort sensefluent to be a subsort of sort fluent. A sensor axiom then is of the form
for a ground set of sensing results . Here is a prime state formula that selects a meaning of the sensing result , whereas the pure prime state formula describes the selected meaning. We stipulate that sensor axioms (which are a form of domain constraint) may only be included if they are modular.
Clearly should be chosen so as to be uniquely determined in each state. If auxiliary axioms are used in then again the programmer must ensure that these evaluate to ground terms in order that a ground state representation can be maintained.
Example 1
The following is the sensor axiom for our goldhunter:
Theoretically, the combination of sensing with the online control of an agent is quite challenging: It is logically sound to to consider the disjunction of all possible sensing results for offline reasoning. In the online setting, however, upon the observation of a sensing result we henceforth have to accept this result as being true; that is, at runtime we add the result to the action theory, something which is logically unsound. On the other hand, it also does not make sense to stipulate that the sensing result be known beforehand.
3.4 Action Theory Representation
We continue by describing how the underlying action theory is represented in ALPprolog. As basic building block we need a representation for prime state formulas. For notational convenience we will represent literals by the (possibly negated) fluent terms only, and, by an abuse of terminology, we will call such a term a fluent literal. A convenient Prolog representation for such a state formula is a list, where each element is either a literal (i.e. a unit clause) or a list of at least two literals (a nonunit clause). In the following we call such a list a PIlist.
Definition 8 (Action Theory Representation)
Action theories as defined in definition 6 are represented in ALPprolog as follows:

The initial state is specified by a Prolog fact
initial_state(PIList).
, wherePIList
mentions only ground fluent literals. Domain constraints other than sensor axioms have to be compiled intoPIList
. 
a Prolog fact
action(A,Precond,EffAx).
, for each action , has to be included, where
A
is an action function symbol, possibly with object terms as arguments; 
Precond
is a PIlist, the action’s precondition; 
EffAx
is a list of cases for the action’s effects with each case being a pairCondEff
, where the effect’s conditionCond
is a PIlist, and the effectsEff
are a list of fluent literals; and 
all variables in
EffAx
also occur inPrecond
.


If present, auxiliary axioms are represented by a set of Prolog clauses. The predicates defined in the auxiliary axioms must be declared explicitly by a fact
aux(Aux).
, whereAux
denotes the listing of the respective predicate symbols.
The sensor axioms are represented as Prolog facts sensor_axiom(s(X),Vals).
, where

s
is a sense fluent with object argumentX
; and 
Vals
is a list ofValIndexMeaning
triples, where
Val
is a pairXresult_i
, whereresult_i
is the observed sensing result; 
Index
is a PIlist consisting of unit clauses; and 
Meaning
is a PIlist, mentioning only fluent literals and only variables fromVal
andIndex
.

The sense fluents have to be declared explicitly by a fact sensors(Sensors).
,
where Sensors
is a listing of the respective function symbols.
This is necessary in order to distinguish sense fluents, ordinary fluents, and auxiliary predicates in PIlists.
3.5 Reasoning for ALPprolog
Reasoning in ALPprolog works as follows: For evaluating the program atoms we readily resort to Prolog. The reasoner for the action theory is based on the principle of progression. Setting out from the initial state, upon each successful evaluation of an action’s precondition against the current state description, we update the current state description by the action’s effects.
Reasoning about the action comes in the following forms:

Given a ground applicable action , from the current state description and the action’s positive and negative effects compute the description of the next state (the update problem).

Given a description of the current state, check whether , where is some state formula in , but not a sense fluent (the entailment problem).

For a sensing action, i.e. a query , integrate the sensing results observed into the agent’s state knowledge (the sensing problem).
In the following we consider each of these reasoning problems in turn.
3.5.1 The Update Problem
It turns out that solving the update problem is very simple. Let State
be a ground PIList, and let Update
be a list of ground fluents. The representation of the next state is then computed in two steps:

First, all prime implicates in
State
that contain either an effect fromUpdate
, or its negation, are deleted, resulting inState1
. 
The next state
NextState
is given by the union ofState1
andUpdate
.
Starting from a ground initial state only ground states are computed.
The correctness of this procedure can be seen e.g. as follows: In [Liu et al. (2006), Drescher et al. (2009)] algorithms for computing updates in a Fluent Calculus based upon Description Logics have been developed. The above update algorithm constitutes a special case of these algorithms.
3.5.2 The Entailment Problem
When evaluating a clause against a ground prime state formula , is first split into the fluent part , and the nonfluent part . It then holds that is entailed by if there is a ground substitution such that

is subsumed by some prime implicate in ; or

some auxiliary atom from can be derived from its defining Prolog clauses.
Computing that the clause is subsumed by can be done as follows:

If is a singleton, then it must be a prime implicate of (modulo unification).

Otherwise there must be a prime implicate in that contains (modulo unification).
Hence the entailment problem for ALPprolog can be solved by
member
, memberchk
, and subset
operations on sorted, duplicatefree lists.
The following example illustrates how reasoning in ALPprolog can be reduced to simple operation on lists. It also illustrates the limited form of reasoning about disjunctive information available in ALPprolog:
Example 2 (Disjunctions and Substitutions in ALPprolog)
Assume that the current state is given by [[at(gold,4),at(gold,5)]]
.
Then the query ?([at(gold,X)])
fails,
because we don’t consider disjunctive substitutions.
However, on the same current state the query ?([[at(gold,X),at(gold,Y)]])
succeeds with X=4
and Y=5
.
3.5.3 The Sensing Problem
Sensing results have to be included into the agent’s state knowledge every time a sensing action is performed,
i.e. a literal ?(s(X))
is evaluated. This works as follows:

First we identify the appropriate sensor axiom
sensor_axiom(s(X),Vals)
. 
Next we identify all the
[Xresult_i]IndexMeaning
triples inVals
such thatresult_i
matches the observed sensing result, and unifyX
withresult_i
. 
We then locate the unique
IndexMeaning
s.t. the current state entailsIndex
. 
Finally, we adjoin
Meaning
to the current state and transform this union to a PIlist.
3.6 Soundness of ALPprolog
At the end of section 3.3 we have already mentioned that adding sensing results to the action theory at runtime makes the subsequent reasoning logically unsound wrt. the original program plus action theory. If we add the set of sensing results observed throughout a run of an ALPprolog program, however, then we can obtain the following soundness result:
Proposition 1 (Soundness of ALPprolog)
Let be a ALPprolog program on top of an action domain . Let be the union of the sensor results observed during a successful derivation of the ALPprolog query with computed answer substitution . Then .
[Proof (Sketch)] It is wellknown that SLDresolution is sound for any ordinary program atom. A query , where is not a sense fluent, is only evaluated successfully if there is a substitution such that . Assume we observe the sensing result for a sense fluent . In general we have (cf. Definition 7):
For soundness, we have to add the observed sensing result as an additional assumption to the theory: .
4 Evaluation
We have evaluated the performance of ALPprolog via the socalled Wumpus World [Russell and Norvig (2003)] that is a wellknown challenge problem in the reasoning about action community. Essentially, the Wumpus World is an extended version of the goldhunter domain from example 1. The main features that make it a good challenge problem are incomplete information in the form of disjunctions and unknown propositions, and reasoning about sensing results.
We have used both Flux and ALPprolog to solve Wumpus Worlds of size up to .^{4}^{4}4The distribution of ALPprolog contains the Wumpus World example for both ALPprolog and Flux. We have done this using three different modellings:

In [Thielscher (2005b)] a Flux model is described that uses quantification over variables — this is beyond ALPprolog.

We have evaluated both languages on a ground model.

We have artificially increased the size of the ground model by making the connections between cells part of the state knowledge.
A first observation is that both languages roughly scale equally well in all models. Using (1) Flux is slightly faster than ALPprolog using (2). Let us then point out that on ground models Flux and ALPprolog maintain the same state representation: Flux also computes the prime implicates. On the encoding (2) ALPprolog is roughly one order of magnitude faster than Flux, whereas on (3) the difference is already two orders of magnitude. The key to the good performance of ALPprolog then is that it handles large state representations well: By encoding states as sorted lists (of lists) some of the search effort necessary in Flux can be avoided. If, however, we use Flux’ capability of handling quantified variables in the state knowledge for a more concise encoding, then ALPprolog and Flux are again on par, with Flux even having slightly the edge. In general, we expect ALPprolog to excel on problem domains that feature large state representations that are not easily compressed using quantification.
It has already been established that Flux gains continuously over standard Golog the more actions have to be performed [Thielscher (2005a)]. As ALPprolog scales as well as Flux the same holds for ALPprolog and Golog. The version of Golog with periodically progressed state knowledge is slightly slower than Flux [Sardina and Vassos (2005)].
Let us also compare ALPprolog, Flux, and Golog from a knowledge representation perspective: Both ALPprolog and Flux allow the programmer to define new auxiliary predicates for the agent strategy that are not present in the action theory, a practically very useful feature that is missing from Golog. Also, the propositional variables used in Golog instead of the finitely many ground terms used in ALPprolog make it hard for the programmer to fully exploit the power of Prolog’s unification mechanism. In this regard Flux, on the other hand, excels in that the programmer can include fluents containing (possibly quantified) variables in the agent’s state knowledge. Contrary to ALPprolog and Golog, however, Flux does not support arbitrary disjunctions.
5 Conclusion and Future Work
In this work we have presented ALPprolog, an efficient logic programming language for the online control of autonomous agents in domains that feature incomplete information and sensing. On the one hand, it can be argued that the stateoftheart languages Golog and Flux already successfully address this application domain. On the other hand, we have shown that ALPprolog excels because of its efficient reasoning with large ground state representations, something that we expect to be quite useful in practice.
For future work, there are two interesting directions: On the one hand it would be nice to extend ALPprolog to offline planning. The disjunctive substitutions in the general ALP proof calculus provide a powerful form of reasoning about conditional plans, or planning in the presence of sensing in the sense of [Levesque (1996)].
On the other hand we plan to fruitfully apply ALPprolog in the domain of General Game Playing. General Game Playing [Genesereth et al. (2005)]
is a new exciting AI research challenge aiming at the integration of manifold AI techniques: A program (also called a player) is given an axiomatisation of the rules of a game. The player then computes a strategy/heuristic that it uses to play and hopefully win the game. The main challenge of General Game Playing consists of constructing suitable heuristics.
However, at its base the player also needs a means to represent, and reason about, the state of the game. Up to now the games played in General Game Playing have been restricted to complete information [Love et al. (2008)] — but clearly games with incomplete information constitute a bigger challenge [Thielscher (2010a)]. We intend to include techniques from ALPprolog into the successful Fluxbased Fluxplayer [Schiffel and Thielscher (2007)].
Acknowledgements.
We appreciate the helpful comments by the reviewers. This work was partially supported by DFG Grant TH 541/14. C. Drescher wishes to acknowledge support by EPSRC Grant EP/G055114/1. M. Thielscher is the recipient of an Australian Research Council Future Fellowship (project number FT 0991348). He is also affiliated with the University of Western Sydney.
References
 Bordini et al. (2007) Bordini, R., Hübner, J., and Wooldridge, M. 2007. Programming MultiAgent Systems in AgentSpeak using Jason. Wiley.
 Drescher et al. (2009) Drescher, C., Liu, H., Baader, F., Guhlemann, S., Petersohn, U., Steinke, P., and Thielscher, M. 2009. Putting abox updates into action. In Proceedings of the Seventh International Symposion on Frontiers of Combining Systems (FroCoS 2009). Trento, Italy.
 Drescher et al. (2009) Drescher, C., Schiffel, S., and Thielscher, M. 2009. A declarative agent programming language based on action theories. In Proceedings of the Seventh International Symposion on Frontiers of Combining Systems (FroCoS 2009). Trento, Italy.
 Genesereth et al. (2005) Genesereth, M. R., Love, N., and Pell, B. 2005. General game playing: Overview of the AAAI competition. AI magazine 26, 2, 62–72.
 Herzig and Varzinczak (2007) Herzig, A. and Varzinczak, I. 2007. Metatheory of actions: Beyond consistency. Artificial Intelligence 171, 16–17, 951–984.
 J.W. Lloyd (1987) J.W. Lloyd. 1987. Foundations of Logic Programming. Springer.
 Levesque et al. (1997) Levesque, H., Reiter, R., Lespérance, Y., Lin, F., and Scherl, R. 1997. GOLOG: A logic programming language for dynamic domains. Journal of Logic Programming 31, 1–3, 59–83.
 Levesque (1996) Levesque, H. J. 1996. What is planning in the presence of sensing? In Proceedings of the Thirteenth National Conference on Artificial Intelligence (AAAI 1996). Portland, Oregon, USA, 1139–1146.
 Liu et al. (2006) Liu, H., Lutz, C., Milicic, M., and Wolter, F. 2006. Updating description logic ABoxes. In Proceedings of the Tenth International Conference on Principles of Knowledge Representation and Reasoning (KR 06). Lake District of the UK.
 Love et al. (2008) Love, N., Hinrichs, T., Haley, D., Schkufza, E., and Genesereth, M. 2008. General game playing: Game description language specification. Tech. rep., Stanford University.
 McCarthy and Hayes (1969) McCarthy, J. and Hayes, P. J. 1969. Some philosophical problems from the standpoint of artificial intelligence. In Machine Intelligence 4, B. Meltzer and D. Michie, Eds. Edinburgh University Press, 463–502.
 Reiter (2001a) Reiter, R. 2001a. Knowledge in Action: Logical Foundations for Describing and Implementing Dynamical Systems. MIT Press, Cambridge, MA.
 Reiter (2001b) Reiter, R. 2001b. On knowledgebased programming with sensing in the situation calculus. ACM Transactions on Computational Logic 2, 4, 433–457.
 Russell and Norvig (2003) Russell, S. J. and Norvig, P. 2003. Artificial Intelligence: a modern approach, 2nd international edition ed. Prentice Hall, Upper Saddle River, N.J.
 Sardina and Vassos (2005) Sardina, S. and Vassos, S. 2005. The wumpus world in IndiGolog: A preliminary report. In Proceedings of the Workshop on Nonmonotonic Reasoning, Action and Change at IJCAI.
 Schiffel and Thielscher (2007) Schiffel, S. and Thielscher, M. 2007. Fluxplayer: A successful general game player. In Proceedings of the Twentysecond National Conference on Artificial Intelligence (AAAI 2007). AAAI Press, Menlo Park, CA, 1191–1196.
 Thielscher (1999) Thielscher, M. 1999. From situation calculus to fluent calculus: State update axioms as a solution to the inferential frame problem. Artificial Intelligence 111, 1–2, 277–299.
 Thielscher (2005a) Thielscher, M. 2005a. FLUX: A logic programming method for reasoning agents. Theory and Practice of Logic Programming 5, 4–5, 533–565.
 Thielscher (2005b) Thielscher, M. 2005b. A FLUX agent for the Wumpus World. In Proceedings of the Workshop on Nonmonotonic Reasoning, Action and Change at IJCAI, L. Morgenstern and M. Pagnucco, Eds. Edinburgh, UK, 104–108.
 Thielscher (2010a) Thielscher, M. 2010a. A general game description language for incomplete information games. In Proceedings of the Twentyfourth National Conference on Artificial Intelligence (AAAI 2010). AAAI Press, Atlanta, 994–999.
 Thielscher (2010b) Thielscher, M. 2010b. Integrating action calculi and AgentSpeak: Closing the gap. In Proceedings of the International Conference on Principles of Knowledge Representation and Reasoning (KR). Toronto, 79–89.
 Thielscher (2011) Thielscher, M. 2011. A unifying action calculus. Artificial Intelligence Journal 175, 120–141.
Comments
There are no comments yet.