Automating Agential Reasoning: Proof-Calculi and Syntactic Decidability for STIT Logics

by   Tim Lyon, et al.
Logic Industries

This work provides proof-search algorithms and automated counter-model extraction for a class of STIT logics. With this, we answer an open problem concerning syntactic decision procedures and cut-free calculi for STIT logics. A new class of cut-free complete labelled sequent calculi G3LdmL^m_n, for multi-agent STIT with at most n-many choices, is introduced. We refine the calculi G3LdmL^m_n through the use of propagation rules and demonstrate the admissibility of their structural rules, resulting in auxiliary calculi Ldm^m_nL. In the single-agent case, we show that the refined calculi Ldm^m_nL derive theorems within a restricted class of (forestlike) sequents, allowing us to provide proof-search algorithms that decide single-agent STIT logics. We prove that the proof-search algorithms are correct and terminate.



There are no comments yet.


page 1

page 2

page 3

page 4


Hypersequent calculi for non-normal modal and deontic logics: Countermodels and optimal complexity

We present some hypersequent calculi for all systems of the classical cu...

Refining Labelled Systems for Modal and Constructive Logics with Applications

This thesis introduces the "method of structural refinement", which serv...

Query-driven PAC-Learning for Reasoning

We consider the problem of learning rules from a data set that support a...

Uniform labelled calculi for preferential conditional logics based on neighbourhood semantics

The preferential conditional logic PCL, introduced by Burgess, and its e...

Cut-free Calculi and Relational Semantics for Temporal STIT Logics

We present cut-free labelled sequent calculi for a central formalism in ...

Syntactic completeness of proper display calculi

A recent strand of research in structural proof theory aims at exploring...

On the proof complexity of logics of bounded branching

We investigate the proof complexity of extended Frege (EF) systems for b...
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

Modal logics of , an acronym for ‘seeing to it that’, have a long tradition in the formal investigation of agency, starting with a series of papers by Belnap and Perloff in the 1980s and culminating in [4]. For the past decades, logic has continued to receive considerable attention, proving itself invaluable in a multitude of fields concerned with formal reasoning about agentive choice making. For example, the framework has been applied to epistemic logic [6], deontic logic [12, 14], and the formal analysis of legal reasoning [6, 13]. Surprisingly, investigations of the mathematical properties of logics are limited [3, 18] and its proof-theory has only been addressed recently [5, 21]. What is more, despite AI-oriented

papers motivating the need of tools for automated reasoning about agentive choice-making

[2, 3, 5], the envisaged automation results are still lacking. The present work will be the first to provide terminating, automated proof-search for a class of logics, including counter-model extraction directly based on failed proof-search.

The sequent calculus [8] is an effective framework for proof-search, suitable for automated deduction procedures. Given the metalogical property of analyticity, a sequent calculus allows for the construction of proofs by merely decomposing the formula in question. In the present work, we employ the labelled sequent calculus—a useful formalism for a large class of modal logics [15, 20]—and introduce labelled sequent calculi (with ) for multi-agent logics containing limited choice axioms, discussed in [22].

In order to appropriate these calculi for automated proof-search, we take up a refinement method presented in [19]—developed for the restricted setting of display logic—and adapt it to the more general setting of labelled calculi. In the refinement process the external character of labelled systems—namely, the explicit presence of the semantic structure—is made internal through the use of alternative, yet equivalent, propagation rules [19]. The tailored propagation rules restrict and simplify the sequential structures needed in derivations, producing, for example, shorter proofs. Moreover, one can show that through the use of propagation rules, the structural rules, capturing the behavior of the logic’s modal operators, are admissible. In our case, the resulting refined calculi derive theorems using only forestlike sequents, allowing us to adapt methods from [19] and provide correct and terminating proof-search algorithms for this class of logics.

In short, the contribution of this paper will be threefold: First, we provide new sound and cut-free complete labelled sequent calculi for all multi-agent logics (with ) discussed in [22]—thus extending the class of logics addressed in [5]. Second, we show how to refine these calculi to obtain new calculi , which are suitable for proof-search. Last, for each , we provide a terminating proof-search algorithm deciding single-agent logic . We conclude by discussing the prospects of generalizing the latter results to the multi-agent setting.

The paper is structured as follows: We start by introducing the class of logics in Sec. 2. In Sec. 3, corresponding labelled calculi are provided, which will subsequently be refined, resulting in the calculi . We devote Sec. 4 to proof-search algorithms and counter-model extraction.

2 Logical Preliminaries

logic refers to a group of modal logics using operators that capture agential choice-making. The basic logic , and the one that will be used throughout this paper, employs two types of modal operators: First, it contains a settledness operator

which expresses which formulae are ‘settled true’ at the current moment. Second, it contains, for each agent

in the language, an atemporal—i.e., instantaneous—choice-operator expressing that ‘agent sees to it that’. This basic choice-operator is referred to as the Chellas [4]. With these two types of operators, one can define a more refined deliberative operator: i.e., iff . Intuitively, holds when ‘agent sees to it that and it is possible for not to hold’. The multi-agent language for is defined accordingly:

Definition 1 (The Language  [11])

Let be a finite set of agent labels s.t. and let be a countable set of propositional variables. is defined via the following BNF grammar:

where and

Notice, the language consists of formulae in negation normal form. This notation allows us to reduce the number of rules in our calculi, enhancing the readability and simplicity of our proof theory. The negation of , written as , is obtained by replacing each operator with its dual and each positive atom with its negation , and each with its positive variant  [5]. Consequently, we obtain the following abbreviations: , , , and . We will freely use these abbreviations throughout this paper. Since we are working in negation normal form, diamond-modalities are introduced as separate primitive operators. We take and as the duals of and , respectively.

Following [11], since we work with instantaneous, atemporal it suffices to regard only single choice-moments in our relational frames. This means that we can forgo the traditional branching time structures of basic, atemporal logic [4]. In what follows, we define frames as those frames in which limits the amount of choices available to each agent to at most -many choices (imposing no limitation when ).111For a discussion of the philosophical utility of reasoning with limited choice see [22].

Definition 2 (Relational Frames and Models)

Let and let for . An -frame is defined as a tuple where is a set of worlds and:

  • For each , is an equivalence relation;

  • For all , ;

  • Let and , then

    For all ,

An -model is a tuple where is an -frame and is a valuation assigning propositional variables to subsets of , i.e. . Additionally, we stipulate that condition (C3) is omitted when .

As in [11], the set of worlds is taken to represent a single moment in which agents from are making their decision. Following (C1), for every agent , the relation is an equivalence relation; that is, functions as a partitioning of into what will be called choice-cells for agent . Each choice-cell represents a set of possible worlds that may be realized by a choice of the agent. The condition (C2) expresses the principle independence of agents, ensuring that any combination of choices, available to different agents, is consistent. The last condition (C3), represents the principle which limits the amount of choices available to an agent to a maximum of . For a philosophical discussion of these principles we refer to [4, Ch. 7C].

Definition 3 (Semantic Clauses for  [5, 11])

Let be an -model and let be a world in its domain . The satisfaction of a formula on at is inductively defined as follows:

  • iff

  • iff

  • iff and

  • iff or

  • iff ,

  • iff ,

  • iff ,

  • iff ,

A formula is globally true on (i.e. ) iff it is satisfied at every world in the domain of . A formula is valid (i.e. ) iff it is globally true on every -model. Last, semantically implies , written , iff for all models and worlds in , if for all , then .

It is worth emphasizing, that the semantic interpretation of refers to the domain of the model in its entirety; i.e. is settled true iff is globally true. This is an immediate consequences of considering instantaneous in a single-moment setting (cf. semantics where a relation is introduced for , e.g. [5]).

The Hilbert calculus for in Fig. 1 is taken from [22]. Apart from the propositional axioms, it consists of axiomatizations for and , for each . It contains the standard bridge axiom , linking to . Furthermore, it contains an independence of agents axiom , as well as a -choice axiom for each . The rules are modus ponens and -necessitation.



Figure 1: The Hilbert calculus for  [4, 22]. A derivation of in from a set of premises , is written as , and is defined inductively in the usual way. When is the empty set, we refer to as a theorem and write .
Theorem 2.1 (Soundness and Completeness [11, 22])

For any formula , if and only if .

3 Refinement of the Calculus

In this section, we introduce labelled calculi for multi-agent STIT logics (with limited choice). Our calculi are modified, extended versions of the labelled calculi for the logics (with ) proposed in [5] and cover a larger class of logics. The calculi possess fundamental proof-theoretic properties such as contraction- and cut-admissibility which follow from the general results on labelled calculi established in [15]. The main goal of this section is to refine the calculi through the elimination of structural rules, resulting in new calculi that derive theorems within a restricted class of sequents. As a result of adopting the approach in [11], the omission of the relational structure corresponding to the modality offers a simpler approach to proving the admissibility of structural rules in the the presence of propagation rules (Sec. 3.2). Let us start by introducing the class of calculi.

3.1 The Calculi

We define labelled sequents via the following BNF grammar:

where , and are from a countable set of labels . Labelled sequents consist exclusively of labelled formulae of the form and relational atoms of the form . For this reason, sequents can be partitioned into two parts: we sometimes use the notation to denote labelled sequents, where is the part consisting of relational atoms and is the part consisting of labelled formulae. Last, we interpret the commas between relational atoms in conjunctively, the comma between and in implicationally, and the commas between labelled formulae in disjunctively (cf. Def.7).

The labelled calculi (where ) are shown in Fig. 2. Note that for each agent , we obtain a copy for each of the rules , , , , and . We refer to , , , and as the structural rules of . The rule captures the independence of agents principle. Furthermore, the rule schema , limiting the amount of choices available to agent , provides different rules depending on the value of in (again, we reserve to assert that the rule does not appear). When , the rule contains premises, where each sequent (for and ) represents a different premise of the rule. As an example, for and the rules for agent are:



Figure 2: The labelled calculi. The superscript on the , , and rule names indicates an eigenvariable condition: the variable occurring in the premise of the rule cannot occur in the context of the premise (or, equivalently, in the conclusion).
Theorem 3.1

The calculi have the following properties:

  1. All sequents of the form are derivable;

  2. Variable-substitution is height-preserving admissible;

  3. All inference rules are height-preserving invertible;

  4. Weakening and contraction are height-preserving admissible:

  5. The cut rule is admissible:

  6. For every formula , is derivable in if and only if , i.e. is sound and complete relative to .


The proof is a basic adaption of [15] and can be found in the App. 0.A.

Proof-theoretic properties like those expressed in (4) and (5) of Thm. 3.1 are essential when designing decidability procedures via proof-search. In constructing a proof of a sequent, proof-search algorithms proceed by applying inference rules of a calculus bottom-up. A bottom-up application of the rule in a proof-search procedure, however, requires one to guess the cut formula , and thus risks non-termination in the algorithm. (One can think of similar arguments why and risk non-termination.) It is thus crucial that such rules are admissible; i.e. everything derivable with these rules, is derivable without them.

Remark 1

To obtain contraction admissibility (Thm. 3.1-(4)) labelled calculi must satisfy the closure condition [15]: if a substitution of variables in a structural rule brings about a duplication of relational atoms in the conclusion, then the calculus must contain another instance of the rule with this duplication contracted.

We observe that if we substitute variable for in the structural rule (below left), we obtain the rule (below right), when the atom is contracted:

Thus, following the closure condition, we must also add to our calculus. However, is a special instance of the rule, and hence it is admissible; therefore, we can omit its inclusion in our calculi. None of the other structural rules possess duplicate relational atoms in their conclusions under a substitution of variables, and so, each calculus satisfies the closure condition.

3.2 Extracting the Calculi

We now refine the calculi, extracting new calculi to which proof-search techniques from [19] may be adapted. In short, we introduce new rules to our calculi, called propagation rules, which are well-suited for proof-search and imply the admissibility of the less suitable structural rules and .

Propagation rules are special sequent rules that possess a nonstandard side condition, consisting of two components. For the first component (1), we transform the sequent occurring in the premise of the rule into an automaton. The labels appearing in the sequent determine the states of the automaton, whereas the relational atoms of the sequent determine the transitions between these states. The following definition, based on [19, Def. 4.1], makes this notion precise:

Definition 4 (Propagation Automaton)

Let be a labelled sequent, be the set of labels occurring in , and . We define a propagation automaton to be the tuple s.t. (i) is the automaton’s alphabet, (ii) is the set of states, (iii) is the initial state, (iv) is the accepting state, and (v) is the transition function where and iff .

We will often write instead of to denote a transition between states. A string is a, possibly empty, concatenation of symbols from (where indicates the empty string). We say that an automaton accepts a string iff there exists a transition sequence from the initial state to the accepting state . Last, we will abuse notation and use equivocally to represent both the automaton and the set of strings accepted by the automaton, i.e. . The use of notation can be determined from the context.

The second component (2) of the rule’s side condition restricts the application of the rule to a particular language that specifies and determines which types of strings occurring in the automaton allow for a correct application of the propagation rule. We define this language accordingly:

Definition 5 (Agent Application Language)

For each , we define the application language to be the language generated from the regular expression , that is, with the empty string.222For further information on regular languages and expressions, consult [17].

Bringing the two components together, a propagation rule is only applied when both the associated propagation automaton accepts a certain string—corresponding to a path of relational atoms in the premise of the rule—and the string is in the application language.

Definition 6 (Propagation Rule)

Let , , and . The propagation rule is defined as follows:

The superscript indicates that for .333Observe that . Hence, deciding which automaton to employ in determining the side condition is inconsequential: when applying the rule top-down we may consult , whereas during bottom-up proof-search we may regard .

We use the represent the set of all propagation rules.

The underlying intuition of the rule (applied bottom-up) is that, given some labelled sequent , a formula is propagated from to another label , if and are connected by a sequence of relational atoms in (with fixed). In the corresponding propagation automaton , this amounts to the existence of a string which represents a sequence of transitions from to , such that all transitions are solely labelled with . To see how the language secures the soundness of the rule, we refer to Thm. 3.3 of App. 0.A. For a general introduction to propagation rules and propagation automata, see [19].

Let us make the introduced notions more concrete by providing an example:

Example 1

Let . The propagation automaton is depicted graphically as (where the single-boxed node designates the initial state and a double-boxed node represents the accepting state):

@/^-1pc/@.¿[rr]—-⟨1 ⟩ u@/^-1pc/@.¿[rr]—-⟨2 ⟩@/^-1pc/@.¿[ll]—-⟨1 ⟩v @/^-1pc/@.¿[rr]—-⟨1 ⟩@/^-1pc/@.¿[ll]—-⟨2 ⟩@/^-1pc/@.¿[ll]—-⟨1 ⟩

Observe that every string the automaton accepts must contain at least one symbol. Since no string of this form exists in , it is not valid to propagate the formula to . That is, the sequent does not follow from applying the propagation rule to the premise.

On the other hand, consider the propagation automaton :

@/^-1pc/@.¿[rr]—-⟨1 ⟩ @/^-1pc/@.¿[rr]—-⟨2 ⟩@/^-1pc/@.¿[ll]—-⟨1 ⟩v @/^-1pc/@.¿[rr]—-⟨1 ⟩@/^-1pc/@.¿[ll]—-⟨2 ⟩z@/^-1pc/@.¿[ll]—-⟨1 ⟩

The automaton accepts the simple string , which is included in the language . Therefore, it is permissible to apply the propagation rule and derive from the premise .

Remark 2

We observe that both of the languages and are regular and, thus, the problem of determining whether , is decidable [19]. Consequently, the propagation rules in may be integrated into our proof-search algorithm without risking non-termination.



Figure 3: The labelled calculus . The superscript on the , , and rules indicate that is an eigenvariable. The side condition is the same as in Def. 6. Last, we have , , and rules for each .

The proof theoretic properties of are preserved when extended with the set of propagation rules (Lem. 1). Moreover, the nature of our propagation rules allows us to prove the admissibility of the structural rules and , for each (resp. Lem. 2 and 3), which results in the refined calculi (shown in Fig. 3). The proofs of Lem. 1 and 2 are present in App. 0.A (the latter is similar to the proof of Lem. 3 presented here).

Lemma 1

The calculi have the following properties: (i) all sequents of the form are derivable; (ii) variable-substitution is height-preserving admissible; (iii) all inference rules are height-preserving invertible; (iv) the , and rules are height-preserving admissible.

Lemma 2 (-Elimination)

Every sequent derivable in is derivable without the use of .

Lemma 3 (-Elimination)

Every sequent derivable in is derivable without the use of .


The result is proven by induction on the height of the given derivation. Also, we evoke Lem. 2 (whose detailed proof is provided in App. 0.A) and assume that all instances of have been eliminated from the given derivation.

Base Case. An application of on an initial sequent (below left) can be re-written as an instance of the rule (below right).

Inductive step. We show the inductive step for the non-trivial cases: and (case (i) and (ii), respectively). All other cases are resolved by applying IH to the premise followed by an application of the corresponding rule.

(i). Let be auxiliary in the inference of the initial derivation (below (1)). Observe that when we apply the rule first (below (2)), the atom is no longer present in , and so, the rule is not necessarily applicable. Nevertheless, we may apply the rule to derive the desired conclusion since . Namely, the fact that only relies on the presence of in .


(ii). Let be the first premise of the initial derivation (below (3)). In the inference of the top derivation, we assume that is auxiliary, that is, the side condition of is satisfied because some string with . (NB. For the non-trivial case, we assume that relies on the presence of , that is, the automaton makes use of transitions or defined relative to .) When we apply the rule first in our derivation (below (4)), we can no longer rely on the relational atom to apply the rule. However, due to the presence of in we may still apply the rule. Namely, since , we know there is a sequence of transitions from to . We replace each occurrence of with and each occurrence of with . There will thus be a string in , and so, the rule may be applied.


As a result of Lem. 13, we obtain cut-free completeness of the calculi :

Theorem 3.2 (Cut-free Completeness of )

For any formula , if , then is cut-free derivable in .


Follows from Thm. 3.1, Lem.’s 13, and the fact that, for each , the rule is admissible, that is, the rule is an instance of the rule .

Last, we must ensure that is sound. To prove this, we need to stipulate how to interpret sequents on -models. Our definition is based on [5]:

Definition 7 (Interpretation, Satisfaction, Validity)

Let be an -model with domain , a labelled sequent, and Lab the set of labels. Let be an interpretation function mapping labels to worlds: i.e. .

is satisfied in with iff for all relational atoms , if holds in , then there must exist some in such that .

is valid iff it is satisfiable in every with any interpretation function .

Theorem 3.3 ( Soundness)

Every sequent derivable in is valid.


We know by Thm. 3.1 that all rules of , with the exception of preserve validity. Details of are given in App. 0.A.

4 Proof-Search and Decidability

In this section, we provide a class of proof-search algorithms, each deciding a logic (with ). (We use to denote the agent in the single-agent setting.) We end the section by commenting on the more complicated multi-agent setting.

4.1 The Single-Agent Setting

In the single-agent case, the independence of agents condition is trivially satisfied, meaning we can omit the rule from each calculus and from consideration during proof-search.

In what follows, we prove that derivations in need only use forestlike sequents. The forestlike structure of a sequent refers to a graph corresponding to the sequent. This control in sequential structure is what allows us to adapt methods from [19] to , and produce a proof-search algorithm that decides , for each . Let us start by making the aforementioned notions precise.

Definition 8 (Sequent Graph)

We define a graph to be a tuple , where is the non-empty set of vertices, the set of edges , and is the labelling function that maps edges from into some non-empty set and vertices from into some non-empty set .

Let be a labelled sequent and let be the set of labels in . The graph of , denoted , is the tuple , where (i) , (ii) iff , and (iii) and iff .

Example 2

The sequent graph corresponding to the labelled sequent = is shown below:

y x[ll]—-1z[ll]—-1
Definition 9 (Tree, Forest, Forestlike Sequent, Choice-tree)

We say that a graph is a tree iff there exists a node , called the root, such that there is exactly one directed path from to any other node in the graph. We say that a graph is a forest iff it consists of a disjoint union of trees.

A sequent is forestlike iff its graph is a forest. We refer to each disjoint tree in the graph of a forestlike sequent as a choice-tree and for any label in , we let represent the choice-tree that belongs to.

The above notions will be significant for proof-search algorithms, for example:

Remark 3

When interpreting a sequent, each choice-tree that occurs in the graph of the sequent is a syntactic representation of an equivalence class of (i.e., a choice-cell for agent ). Using this insight, we know that if agent is restricted to -many choices, then if there are choice-trees in the sequent, at least two choice-trees must correspond to the same equivalence class in . We use this observation to specify how is applied in the algorithm.

The following definitions introduce the necessary tools for the algorithms:

Definition 10 (Saturation, -realization, , -propagated)

Let be a forestlike sequent and let w be a label in .

The label is saturated iff the following hold: (i) If , then , (ii) if , then and , (iii) if , then or .

A label in is -realized iff for every , there exists a label such that . A label in is -realized iff for every