Towards a Formal Theory of Communication for Multiagent ... - IJCAI
Towards a Formal T h e o r y of C o m m u n i c a t i o n f o r M u l t i a g e n t Systems M u n i n d a r P. S i n g h * Dept of Computer Sciences and Artificial Intelligence Lab University of Texas MCC Austin, TX 78712-1188 Austin, TX 78759 USA USA Abstract Agents in multiagent systems interact to a large extent by communicating. Such communicat i o n may be f r u i t f u l l y studied f r o m the point of view of speech act theory. In order for multiagent systems to be formally and rigorously designed and analyzed, a semantics of speech acts t h a t gives their objective modeltheoretic conditions of satisfaction is needed. However, most research into multiagent systems t h a t deals w i t h communication provides only i n f o r m a l descriptions of the different message types used. A n d this problem is not addressed at all by t r a d i t i o n a l speech act theory or by AI research i n t o discourse understanding. I provide a formal semantics for the maj o r kinds of speech acts at a level t h a t has not been considered before. T h e resulting theory applies u n i f o r m l y to a wide range of multiagent systems. Some applications of this theory are o u t l i n e d , and some of its theorems listed.
1
Introduction
M u l t i a g e n t systems are ubiquitous in A r t i f i c i a l I n t e l l i gence. Even in the simplest such systems, the epithet is justified only if the agents involved interact w i t h each other in different ways. One of the most n a t u r a l ways in which intelligent interaction may occur is through c o m m u n i c a t i o n , especially communication about action. Agents may c o m m a n d , request, advise, or p e r m i t each other to do certain actions. They may also promise actions of their o w n , or p r o h i b i t those of others. When complex multiagent systems are to be designed or analyzed a formal theory of the kinds of communication that may take place among agents would be crucial. Unfortunately, no theory is currently available that provides the objective semantics of the messages exchanged. T h i s paper describes research t h a t has been done to fill this v o i d . T h i s research uses ideas about the a b i l i t y and intentions of situated agents t h a t were motivated and *This research was partially supported by the Microelectronics and Computer Technology Corporation, and by the National Science Foundation (through grant # IRI-8945845 to the Center for Cognitive Science, University of Texas).
developed on independent grounds, albeit w i t h a view to their final application to this problem [Singh, 1991; Singh, 1990a; Singh, 1990b]. T h i s connection to other theories is reason to be reassured t h a t this theory is not ad hoc, and w i l l coherently fit in a bigger picture. The theory presented in this paper has ramifications in several subareas of A I , notably, multiagent planning and act i o n , autonomous agents, and cooperative work [Gasser and Huhns, 1989; Huhns, 1987]. T r a d i t i o n a l l y , speech act theory classifies communications or messages into several kinds of illocutionary acts [Searle, 1969; Searle and Vanderveken, 1985]. These include assertivcs, directives, commisives, permissives and prohibitives. Briefly, assertives are statements of fact; directives are commands, requests or advice; commisives (e.g., promises) c o m m i t the speaker to a course of action; permissives issue permissions; and prohibitives take them away. Classical logic applies only to the case of assertivcs and considers only their t r u t h and falsity. Therefore, it is inappropriate for other kinds of speech acts ( H a m b l i n describes and criticizes several nonclassical logics for commands [ H a m b l i n , 1987, pp. 97-136], so I do not consider them here). Research in speech act theory, on the other h a n d , concentrates on describing the conditions under which a particular speech act (of whatever form) may be said to have occurred [Grice, 1969; Searle, 1975]. T h e AI l i t e r a t u re in this area too is concerned w i t h the linguistic or discourse-related aspects of this problem (e.g., for identifying the illocutionary force of indirect speech acts [Allen and Perrault, 1980], or defining their effects on the m u t u a l beliefs of agents [Cohen and Levesque, 1988]). Of interest here is the orthogonal problem of formally describing the conditions of satisfaction for the different kinds of speech acts. I take the view t h a t communicat i o n occurs because agents need to interact effectively and to influence each others' actions. W h i l e the illocutionary force of a speech act can be t r i v i a l l y determined f r o m the syntax (in an artificial language that our agents would use), the objective conditions of the satisfaction of a speech act are a part of the semantics. A formal semantics is i m p o r t a n t because (1) as designers and analyzers, we need a rigorous understanding of communication in the systems we design; and (2) we would often like to embed a version of the semantics in the agents themselves so they can use it in their reasoning about their
Singh
69
own (and others') speech acts. T h e m a i n original contributions of this paper are the following: (1) It argues t h a t there is a level of f o r m a l semantics of speech acts t h a t is distinct f r o m b o t h (a) what is t r a d i t i o n a l l y considered their semantics, namely, the conditions under which they may be said to have occurred; and (b) their pragmatics, namely, the effects they may or ought to have on the speaker's and hearer's cognitive states. I.e., the proposed semantics differs f r o m b o t h the illocutionary and the perlocutionary aspects of speech acts. (2) T h i s paper argues t h a t the semantics of speech acts roughly corresponds to the conditions under which we would affirm t h a t the given speech act had been satisfied. (3) It is proposed t h a t this semantics can be captured in the usual model-theoretic framework by i n t r o d u c i n g different operators t h a t distinguish the satisfaction of a speech act f r o m its mere occurrence. (4) T h e actual definitions are to be given in terms of the intentions and know-how of the participants and the state of the w o r l d (at some salient t i m e or times). A p r o b l e m n o t addressed here concerns the effects a speech act has on the hearer. These depend on issues like the social relationship of the agents or on matters of performance—these are not easy to describe, and are connected to processes of deliberation and belief revision [Perrault, 1987], rather t h a n to the semantics of communication per se. Perrault provides some postulates for such revision using default logic. His focus is on the pragmatics of speech acts in n a t u r a l language understanding, rather t h a n the semantics as considered here. In any case, a semantics would help clarify our intuitions even about the pragmatic aspects of speech acts. As a clarification of my goals, note t h a t the role of the proposed semantics is akin to t h a t of classical semantics for assertives. Classical semantics only tells us when an assertive is objectively satisfied—it makes no claims about when an assertive should actually be uttered or believed. Therefore, the focus here is on satisfaction conditions. T h e conditions of satisfaction for most kinds of speech acts differ significantly f r o m those of assertives t h a t are o r d i n a r i l y considered in logic. Assertives, being claims of fact, are true or false; other speech acts call for a more complex framework in which their felicity or success can be described. In the context of imperatives, Haniblin distinguishes between w h a t he calls exiensional and whole-hearted satisfaction [ H a m b l i n , 1987, p p . 153157]. Briefly, the former n o t i o n admits accidental success, while the latter does n o t . H a m b l i n realized t h a t these were useful things t h a t may be said of a speech act; however, his a i m was simply to be able state prescriptive conditions on when w h a t k i n d of imperatives ought to be issued, and the philosophical problems t h a t arise when one is in a "quandary." T h u s his focus seems to have been p r a g m a t i c . I take advantage of some of his ideas, b u t make a finer distinction and extend it to other i m p o r t a n t kinds of speech acts here, f o r m a l l y rel a t i n g t h e m to intentions and know-how in the process. In §4.2, I address the problem of w h a t k i n d of prescriptive constraints on communication may be stated, b u t see t h a t as essentially supervenient on the semantics. In §2, I discuss three different senses of satisfaction
70
Architectures and Languages
for the five kinds of speech acts considered in this paper. In §3, I describe a theory of intentions and know-how, and a f o r m a l model which I then use to formalize the different notions of satisfaction. In §4, I show how this theory may be used in the design of multiagent systems, and list some useful theorems.
2
Shades of S a t i s f a c t i o n
As remarked above, c o m m u n i c a t i o n among agents in a multiagent system can be best understood by appealing to speech act theory [ A u s t i n , 1962; Searle, 1969]. In speech act theory an " i l l o c u t i o n " (which I identify w i t h a message) is seen to have two parts: an illocutionary force and a proposition. T h e i l l o c u t i o n a r y force distinguishes, e.g., a command f r o m a promise; the proposit i o n describes the state of the w o r l d t h a t is, respectively, commanded or promised. T h i s suggests a simple syntax for messages in an artificial language. A message, m, is a pair ( i , p ) , where i identifies the illocutionary force, and p the proposition. Here i is an atomic symbol f r o m the set {directive, commisive, permissive, proh i b i t i v e , assertive}; and p is a logical f o r m u l a . T h i s much is quite standard even in the AI literature t h a t deals w i t h communication among agents [Huhns et al., 1990; T h o m a s et al., 1990]. However, none of the AI papers so far give a rigorous f o r m a l semantics for messages of different i l l o c u t i o n a r y forces. T h i s lacuna is filled by this paper. B u t before I come to the formalizat i o n , I must discuss the different senses of satisfaction of speech acts. T h e rest of this section extends the discussion in [ H a m b l i n , 1987, pp. 153-157]. Note t h a t these different senses agree for the case of assertives. T h e propositional part of a message specifies the state of the world t h a t the message is, in some sense, about. An assertive asserts of t h a t state t h a t it holds (i.e., currently, though the proposition could be temporally indexed); a directive asks the hearer to b r i n g t h a t state about; a commisive commits the speaker to bringing it about; and so o n . T h u s the satisfaction of a message depends b o t h on its i l l o c u t i o n a r y force and its proposition. T h e different notions of satisfaction are motivated using directives; other speech acts are considered in §3. In the simplest sense of satisfaction, called exiensional satisfaction in [ H a m b l i n , 1987, p. 153], a message is said to be satisfied ( w i t h only m i n o r qualifications) j u s t if its proposition turns out to h o l d . E.g., a directive is satisfied when the proposition becomes true. Extensional satisfaction looks only at the i m m e d i a t e state of the w o r l d — the proposition may have been made true accidentally, or for the " w r o n g reasons," b u t would s t i l l meet the requirements for extensional satisfaction. T h i s notion of satisfaction meets weak behavioral specifications; e.g., if the success of a speech act is p a r t of a p l a n , then when it succeeds the agent can legitimately proceed to the next stage of the plan. However, this n o t i o n is not acceptable for complex systems because f o r t u i t o u s circumstances would n o t , in general, arise often in t h e m . A system whose agents were designed on this basis would t u r n o u t to be not sufficiently robust—e.g., we would expect a guarantee t h a t some directives would be satisfied in a variety of circumstances, rather t h a n t h a t they once
were. Also, it is of not much help practically, since it yields no insights about how the i n d i v i d u a l agents in a multiagent system ought to be designed. T h i s motivates the next sense of satisfaction, wholehearted satisfaction. T h e whole-hearted satisfaction of a directive requires not only t h a t the specified proposit i o n be made t r u e , b u t be made true in a sure-fire manner. T h e concerned agent should not only bring about the r i g h t state of the w o r l d , b u t know how to bring it about and intend to b r i n g it about (thus it would bring it about in a way t h a t exploits its know-how). By using whole-hearted satisfaction, the designer can require t h a t an agent not issue two commands, which cannot b o t h be whole-heartedly satisfied (due to l i m i t e d know-how), even if they can b o t h be extensionally satisfied. Even whole-hearted satisfaction admits cases where the relevant proposition was made true, b u t only because it was going to be made true anyway, irrespective of whether the given speech act was performed. I.e., the speech act was pointless and played no real role in its satisfaction. T h i s happens when, for a directive, the bearer was going to do the desired actions anyway. Often, it is useful to eliminate these conditions, so that the given speech act is really necessary. T h i s requires that not only must the proposition in the speech act be made true in a sure-fire way, it must be made true because of t h a t speech act. T h i s , relevant satisfaction, is the strongest notion of satisfaction t h a t 1 consider here. T h e taxonomy of speech acts of this paper is motivated by the fact t h a t permissives, prohibitives and directives have different satisfaction conditions (cf. [Bach and II arnish, 1979, pp. 39-54] and [Searle and Vanderveken, 1985, ch. 9], where permissives and prohibitives are lumped together w i t h directives). T h e more conventionoriented or culture-oriented illocutionary forces (e.g., christenings, greetings) are not considered here. Interrogatives are semantically quite like directives, b u t need special treatment to allow for answers; they are not i n cluded in this paper for reasons of space.
3
F o r m a l i z a t i o n of S a t i s f a c t i o n Conditions
T h e different notions of satisfaction of speech acts depend on the definitions of know-how and intention. A formal rigorous theory of situated know-how and intentions has already been developed and reported. Here the same framework and technical definitions are used to give an account of the different sorts of satisfaction of several kinds of speech acts. T h i s project has been inspired by t h a t of C. L. H a m b l i n , who died while w r i t ing [ H a m b l i n , 1987]. He argued t h a t an account of i m peratives must be b u i l t on top of a theory of abilities and intentions, especially, one which is " . . . n o t hidden in the m i n d , however, b u t expressed in action . . . " [Hamb l i n , 1987, foreword by Belnap, p. v i i i ] . He d i d not have such a f o r m a l theory of know-how and intentions, b u t I take this advice seriously, and t r y to give a semantics of several kinds of speech acts in terms of my earlier theory of intentions and know-how, where these concepts are defined in terms of the actions of agents situ-
ated in an objective model [Singh, 1990b; Singh, 1990a; Singh, 1991]. I briefly describe that theory, and then t u r n to the formal model. 3.1
Know-how and Intentions
T h i s theory of know-how and intentions is meant to apply to b o t h t r a d i t i o n a l plan-based architectures and modern situated ones. It uses the concept of strategies as abstract descriptions of the agents' behavior. Strategies correspond to plans in t r a d i t i o n a l systems and to the architectural structure of reactive agents, as instantiated at a given t i m e . A strategy is simply the designer's description of the agent and the way in which it behaves. An agent knows how to achieve A, if it can achieve A whenever it so "intends." Let each agent have a strategy t h a t it follows in the current situation. Intuitively, an agent knows how to achieve A relative to a strategy Y, iff it possesses the skills required to follow Y in such a way as to achieve A. Thus know-how is partitioned into two components: the " a b i l i t y " to have satisfactory strategies, and the " a b i l i t y " to follow them. Similarly, an agent intends to achieve A by performing a strategy Y, iff it currently has strategy Y and the successful performance of Y by it entails A; i.e., iff the agent is t r y i n g to perform Y to achieve A. Note t h a t having an intention does not entail having the know-how to match i t , and vice versa. Another i m p o r t a n t p r i m i t i v e is can-prevent, notated Kprev. T h i s is related to know-how and applies when the given agent is able to perform actions so as to prevent the occurrence of the given condition. For reasons of space, the technical details of [Singh, 1991; Singh, 1990a; Singh, 1990b] are not included here. The presentation below is self-contained, however. 3.2
The Formal Model
T h e formal model here is based on possible worlds. Each possible world has a branching history of times. Agents influence the future by acting, but the outcome also depends on other events. A scenario at a given world and time is any linear branch of the f u t u r e beginning there— this corresponds to a particular run or trace of the given system. A subscenario is a t r i p l e , ( S , t , t ' ) , which denotes a section of scenario S f r o m t i m e t to t'. T h e interpretation [] assigns sets of world-times pairs to predicates, for each possible tuple of their arguments; it assigns sets of subscenarios to each action, for each agent who might do i t . T r u t h in the model, M, is defined relative to a world w and a t i m e t: M \=w,t P denotes t h a t p is true in M at w and t. Another useful notion is of t r u t h relative to a scenario and a t i m e in it (a scenario determines the world): M \=s,t P denotes t h a t p is true in M on S at t. T h e formal language here is the predicate calculus, augmented w i t h temporal logic (used in §4.3), and three predicates, 'K h o w , 'intends' and 'K'pre,,,' each applying to an agent, a strategy and a f o r m u l a . Speech acts are, first of a l l , actions. For simplicity, they are seen as the actions of j u s t their speakers, and occur over subscenarios. Let 'says-to' be a parametrized speech act, to be used as 'says-to(y, m).' T h i s action w i l l be seen as an action done by agent x. (ts,,th) E [ s a y s - t o ( y , m ) ] x means t h a t agent x commu-
Singh
71
nicated message m to agent y in the t i m e f r o m t s (the time of speaking) to t h (the t i m e of hearing). T h i s j u s t says t h a t the message was successfully t r a n s m i t t e d ( i t is possible to allow failed transmissions, b u t t h a t is not useful here). Define a new predicate ' c o m m ' t h a t applies to two agents, and a message. ' C o m m ( x , y , m ) ' is true at w, t j u s t if x said (or started to say) rn to y then. A t r a n s m i t t e d message may, of course, not be satisfiable. In order to be able to t a l k of the different kinds of satisfaction of messages, I introduce three operators: ESAT, WSAT and RSAT (collectively called SAT below) t h a t apply on formulas of the f o r m ' c o m m ( i , y , m ) . ' They respectively state t h a t the given message is extensionally, whole-heartedly, and relevantly satisfied. Now the conditions of t r u t h are given for each of these three operators and for each of the possible illocutionary forces, relative to a scenario and a t i m e .
3.3
Extensional Satisfaction
Extensional satisfaction is defined relative to a scenario and a time in it (thus " f u t u r e " means future w i t h i n t h a t scenario). A directive is satisfied at a scenario and time j u s t if its proposition becomes true at a future time on t h a t scenario. A commisive too is satisfied j u s t if its proposition becomes true at some future t i m e on the given scenario. An assertive is satisfied j u s t at those times where its proposition is true. A permissive is satisfied on a scenario and t i m e j u s t if it is taken advantage of sometimes in the f u t u r e of t h a t t i m e on t h a t scenario. A p r o h i b i t i v e is satisfied j u s t if it is never violated in the f u t u r e of the given t i m e on the given scenario. T h u s the agents' intentions and know-how do not m a t t e r for extensional satisfaction.
3.4
Whole-hearted Satisfaction
Whole-hearted satisfaction too is defined relative to a scenario and a t i m e . A directive is satisfied on a scenario
72
and t i m e j u s t if its proposition becomes true at a future t i m e in t h a t scenario, and all along the scenario f r o m the given t i m e to t h e n , the hearer has the know-how, as well as the i n t e n t i o n to achieve i t . I.e., if the hearer has a strategy ( t h a t it may be said to be following) relative to which it has the know-how and the intention to achieve p (as explained in §3.1). Similarly, a commisive is satisfied j u s t when its proposition becomes true at some f u t u r e t i m e on the given scenario, and all along the scenario f r o m the given t i m e to t h e n , the speaker has the know-how to achieve it and also intends i t . T h e condition for assertives is unchanged. A permissive is satisfied at a scenario and a t i m e j u s t if it is taken advantage of by the hearer at a f u t u r e point on t h a t scenario. B u t when a permissive is taken advantage of, it allows the hearer to do actions at certain times t h a t it could not have done before because they m i g h t possibly lead to the c o n d i t i o n becoming true. Thus a permissive is satisfied on a scenario on which the hearer does at least one action whose performance can lead to a situation where it is unable to prevent t h a t condition f r o m occurring (i.e., the hearer can now risk l e t t i n g t h a t condition hold). Similarly, a p r o h i b i t i v e is satisfied at a scenario and t i m e j u s t if none of the actions done by the hearer on t h a t scenario ( i n the f u t u r e ) , can lead to a situation where the hearer would be unable to prevent the condition f r o m occurring (i.e., the hearer does not risk v i o l a t i n g the p r o h i b i t i o n ) .
3.5
Relevant Satisfaction
Relevant satisfaction is also defined relative to a scenario and a t i m e . It resembles the previous case, b u t differs in adding a requirement t h a t roughly says t h a t the given speech act is the true reason for its success. A directive is satisfied j u s t when its proposition becomes true at a t i m e in the f u t u r e of the given t i m e , and all along the scenario f r o m the given t i m e to t h e n , the hearer has the know-how to achieve i t , and furthermore t h a t this know-how does not arise in at least one scenario t h a t is a t e m p o r a l alternative to the given one. T h e definition for commisives is analogous. T h e condition for assertives continues to be the same as before. A permissive is satisfied at a scenario and a t i m e j u s t if it is taken advantage of by the hearer at a f u t u r e t i m e on the given scenario, w i t h an a d d i t i o n a l requirement t h a t
4
Applying the Theory
T h e two m a i n m o t i v a t i o n s for developing this theory are to provide a rigorous f o u n d a t i o n for the design of m u l tiagent systems and to j u s t i f y some prescriptive claims about how agents should communicate in t h e m . T h i s theory has given objective criteria to evaluate the correctness of different scenarios, or runs of a multiagent system. In design, the problem is to create a system which allows only correct scenarios to be actualized. Prescriptive claims for agents tell them what to do given their beliefs and intentions, so t h a t correct scenarios emerge. I discuss these two problems below, and then some i m p o r t a n t f o r m a l consequences of this theory. 4.1
Designing M u l t i a g e n t Systems
have (at the appropriate times) (1) the know-how to answer correctly whether they w i l l b i d on a given task, and to promise as described above; and (2) the corresponding intentions. Thus the design goal reduces to ensuring conditions (1) and (2) for all the agents. T h e different senses of SAT yield rigorous definitions for three kinds of correctness conditions for multiagent systems. A scenario can be defined to be correct in the sense of SAT if all the messages passed on it are satisfied in the same sense. In general terms, the designer's goal is to ensure t h a t all runs t h a t may be actualized are correct. T h i s reduces to the goal t h a t the intentions and know-hows of the agents are such t h a t only correct scenarios are actualized. T h i s is the sense of correctness t h a t designers use in practice. They usually achieve this k i n d of correctness by, e.g., h a r d - w i r i n g the intention to cooperate in their agents, or by setting up hierarchical structures such t h a t some directives (commands) are always obeyed, and others (requests) obeyed whenever they do not conflict w i t h the hearer's current intentions. 4.2
Normative Constraints on Communication
T h e objective criteria given above can be used to motivate some normative constraints on communication among the agents in a multiagent system. These constraints could be used by the designer, and possibly by the agents themselves to reason about the messages they exchange. Note t h a t these are j u s t meant as weak constraints, and may be easily overridden. If (some of) these constraints are obeyed, the scenarios t h a t are actualized are not j u s t correct b u t also " g o o d . " Imposing these constraints can simplify a design since certain good properties can then be taken for granted. 1. An agent should issue a directive only if its intentions are satisfied (i.e., its current strategies arc performed successfully) on all scenarios on which the directive is W S A T . T h i s is the converse of the defi n i t i o n for M )=w,t WSATq suggested in §3.2—this ensures t h a t actions by the hearer w i l l not render the speaker's own intentions impossible to achieve.
One extension needed is to distinguish between messages of the same m a j o r illocutionary force t h a t are different in some i m p o r t a n t respects. E.g., commands differ f r o m requests, since they presuppose a u t h o r i t y on part of the speaker, and can cause a change in the hearer's intentions under a wider variety of circumstances t h a n requests. A correct scenario w i t h a command must WSAT it as a directive (assuming a u t h o r i t y ) ; one w i t h a request must WSAT it t o o , b u t only if some other conditions on cooperation are met. T h e designer needs to constrain the issuing of directives, a n d / o r increase know-how and constrain intentions so t h a t the system actualizes only correct scenarios. E.g., requiring t h a t agents persist w i t h their strategies for sufficiently long considerably s i m p l i fies generating runs t h a t on which directives and commisives are W S A T . Correctness must be ensured for all the message types t h a t can occur in the system.
3. Agents ought to persist w i t h their strategies (i.e., their strategy at a later t i m e should be the appropriate " t a i l " of their strategy at an earlier t i m e ) . T h i s not only simplifies the W S A T and RSAT conditions for directives and commisives, b u t also s i m p l i fies the interactions among agents by, e.g., m a k i n g it easy for an issuer of a directive to not take on a strategy t h a t would interfere w i t h its compliance.
As an example, consider the contract net [Davis and S m i t h , 1983]. In its simplest f o r m , a manager sends out a call for bids to all contractors—treat this as requesting the hearer to b i d and to promise t h a t it w i l l do the task for a certain price ( i f it wants to b i d ) . Contractors who b i d thus promise t h a t if requested to do the task and given their price, they w i l l do i t . T h e manager selects one contractor and requests it to do the task. In any correct design for this protocol, the contractors must
4. A l l messages sent by one speaker must be m u t u ally consistent in the sense of being j o i n t l y satisfiable on at least some scenarios. E.g., different d i rectives should not clash w i t h each other and prohibitives should not preclude the satisfaction of d i rectives and other prohibitives. T h i s prevents many unacceptable situations, b u t can cause problems if some redundant permissions are issued, i.e., those t h a t m i g h t never be used on some scenarios where
2. A l l messages sent must be RSAT, where the quantification over scenarios is restricted to scenarios compatible w i t h the speaker's current intentions.
Singh
73
References [Allen and Perrault, 1980] J. F. A l l e n and C. R. Perrault. A n a l y z i n g intention in utterances. Artificial Intelligence, 15:143-178, 1980. [ A u s t i n , 1962] J. L. A u s t i n . How to do Things with Words. Clarendon, O x f o r d , U K , 1962. [Bach and Harnish, 1979] K. Bach and R. M. H arnisn. Linguistic Communication and Speech Acts. MIT Press, Cambridge, M A , 1979. [Cohen and Levesque, 1988] P. R. Cohen and H. J. Levesque. R a t i o n a l interaction as the basis for comm u n i c a t i o n . Tech. Rep. 433, S R I I n t e r n a t i o n a l , Menlo Park, C A , A p r . 1988. [Davis and S m i t h , 1983] R. Davis and R. G. S m i t h . Neg o t i a t i o n as a metaphor for distributed problem solving. Artificial Intelligence, 20:63-109, 1983. [Emerson, 1989] E. A. Emerson. Temporal and modal logic. In J. van Leeuwen, ed., Handbook of Theoretical Computer Science. N o r t h - H o l l a n d Publishing Company, A m s t e r d a m , 1989. [Gasser and Huhns, 1989] L. Gasser and M. N. Huhns, eds. Distributed Artificial Intelligence, Volume II. P i t man, London,1989. [Grice, 1969] P. Grice. Utterer's meaning and intentions. Philosophical Review, 1969. [ H a m b l i n , 1987] C. L. H a m b l i n . Imperatives. Blackwell L t d . , O x f o r d , U K , 1987.
Basil
[Huhns, 1987] M. N. Huhns, ed. Distributed Artificial Intelligence. P i t m a n , L o n d o n , 1987.
5
Conclusions
None of the extant theories of communication address quite the same problem as I have addressed in this paper. T h e theory presented here refines and formalizes some i n t u i t i o n s about communication among agents. 1 started o u t w i t h a set of obvious and well-known i n t u itions about the nature of communication derived f r o m classical speech act theory. However, I motivated and approached an i m p o r t a n t problem t h a t has not been addressed in the literature on speech act theory, or even in the AI literature. My c o m m i t m e n t is stronger to the general claims t h a n to the specifics. However, using definitions of the intentions and know-how of an agent, I was systematically able to give rigorous definitions of the conditions of satisfaction for speech acts of different of i l l o c u t i o n a r y forces. These definitions capture many of our i n t u i t i o n s about when, as speakers and hearers, we believe t h a t a given speech act has been satisfied. T h i s theory has applications in the design of multiagent systems, where constraints on the know-how and intentions of agents are derived f r o m their desired communicative behavior. It can also yield some well-motivated normative constraints on communication among agents. An advantage of the f o r m a l approach is t h a t this process can be guided by the several theorems t h a t exist.
74
Architectures and Languages
[Huhns et al., 1990] M. N. Huhns, D. Bridgeland, and N. A m i . A D A I communication aide. Tech. Rep. A C T RA-317-90, Microelectronics and Computer Technology C o r p . , A u s t i n , T X , Oct. 1990. [Perrault, 1987] R. Perrault. An application of default logic to speech act theory. Tech. Rep. 90, C S L I , Stanf o r d , C A , Mar. 1987. [Searle and Vanderveken, 1985] J. R. Searle and D. Vanderveken. Foundations of Illocutionary Logic. Cambridge University Press, Cambridge, U K , 1985. [Searle, 1969] J. R. Searle. Speech Acts. Cambridge U n i versity Press, Cambridge, U K , 1969. [Searle, 1975] J. R. Searle. Indirect speech acts. In P. Cole and J. L. M o r g a n , eds., Syntax and Semantics, Volume 8. Academic Press, New Y o r k , N Y , 1975. [Singh, 1990a] M. P. Singh. G r o u p intentions. In 10th Workshop on Distributed AI, Oct. 1990. [Singh, 1990b] M. P. Singh. Towards a theory of situated know-how. In ECAI, A u g . 1990. [Singh, 1991] M. P. Singh. A logic of situated know-how. In AAAI, J u l y 1991. [Thomas et al., 1990] B. T h o m a s , Y. Shoham, and A. Schwartz. Modalities in agent-oriented p r o g r a m m i n g . C o m p u t e r Science D e p a r t m e n t , Stanford University, Feb. 1990.
1
Introduction
M u l t i a g e n t systems are ubiquitous in A r t i f i c i a l I n t e l l i gence. Even in the simplest such systems, the epithet is justified only if the agents involved interact w i t h each other in different ways. One of the most n a t u r a l ways in which intelligent interaction may occur is through c o m m u n i c a t i o n , especially communication about action. Agents may c o m m a n d , request, advise, or p e r m i t each other to do certain actions. They may also promise actions of their o w n , or p r o h i b i t those of others. When complex multiagent systems are to be designed or analyzed a formal theory of the kinds of communication that may take place among agents would be crucial. Unfortunately, no theory is currently available that provides the objective semantics of the messages exchanged. T h i s paper describes research t h a t has been done to fill this v o i d . T h i s research uses ideas about the a b i l i t y and intentions of situated agents t h a t were motivated and *This research was partially supported by the Microelectronics and Computer Technology Corporation, and by the National Science Foundation (through grant # IRI-8945845 to the Center for Cognitive Science, University of Texas).
developed on independent grounds, albeit w i t h a view to their final application to this problem [Singh, 1991; Singh, 1990a; Singh, 1990b]. T h i s connection to other theories is reason to be reassured t h a t this theory is not ad hoc, and w i l l coherently fit in a bigger picture. The theory presented in this paper has ramifications in several subareas of A I , notably, multiagent planning and act i o n , autonomous agents, and cooperative work [Gasser and Huhns, 1989; Huhns, 1987]. T r a d i t i o n a l l y , speech act theory classifies communications or messages into several kinds of illocutionary acts [Searle, 1969; Searle and Vanderveken, 1985]. These include assertivcs, directives, commisives, permissives and prohibitives. Briefly, assertives are statements of fact; directives are commands, requests or advice; commisives (e.g., promises) c o m m i t the speaker to a course of action; permissives issue permissions; and prohibitives take them away. Classical logic applies only to the case of assertivcs and considers only their t r u t h and falsity. Therefore, it is inappropriate for other kinds of speech acts ( H a m b l i n describes and criticizes several nonclassical logics for commands [ H a m b l i n , 1987, pp. 97-136], so I do not consider them here). Research in speech act theory, on the other h a n d , concentrates on describing the conditions under which a particular speech act (of whatever form) may be said to have occurred [Grice, 1969; Searle, 1975]. T h e AI l i t e r a t u re in this area too is concerned w i t h the linguistic or discourse-related aspects of this problem (e.g., for identifying the illocutionary force of indirect speech acts [Allen and Perrault, 1980], or defining their effects on the m u t u a l beliefs of agents [Cohen and Levesque, 1988]). Of interest here is the orthogonal problem of formally describing the conditions of satisfaction for the different kinds of speech acts. I take the view t h a t communicat i o n occurs because agents need to interact effectively and to influence each others' actions. W h i l e the illocutionary force of a speech act can be t r i v i a l l y determined f r o m the syntax (in an artificial language that our agents would use), the objective conditions of the satisfaction of a speech act are a part of the semantics. A formal semantics is i m p o r t a n t because (1) as designers and analyzers, we need a rigorous understanding of communication in the systems we design; and (2) we would often like to embed a version of the semantics in the agents themselves so they can use it in their reasoning about their
Singh
69
own (and others') speech acts. T h e m a i n original contributions of this paper are the following: (1) It argues t h a t there is a level of f o r m a l semantics of speech acts t h a t is distinct f r o m b o t h (a) what is t r a d i t i o n a l l y considered their semantics, namely, the conditions under which they may be said to have occurred; and (b) their pragmatics, namely, the effects they may or ought to have on the speaker's and hearer's cognitive states. I.e., the proposed semantics differs f r o m b o t h the illocutionary and the perlocutionary aspects of speech acts. (2) T h i s paper argues t h a t the semantics of speech acts roughly corresponds to the conditions under which we would affirm t h a t the given speech act had been satisfied. (3) It is proposed t h a t this semantics can be captured in the usual model-theoretic framework by i n t r o d u c i n g different operators t h a t distinguish the satisfaction of a speech act f r o m its mere occurrence. (4) T h e actual definitions are to be given in terms of the intentions and know-how of the participants and the state of the w o r l d (at some salient t i m e or times). A p r o b l e m n o t addressed here concerns the effects a speech act has on the hearer. These depend on issues like the social relationship of the agents or on matters of performance—these are not easy to describe, and are connected to processes of deliberation and belief revision [Perrault, 1987], rather t h a n to the semantics of communication per se. Perrault provides some postulates for such revision using default logic. His focus is on the pragmatics of speech acts in n a t u r a l language understanding, rather t h a n the semantics as considered here. In any case, a semantics would help clarify our intuitions even about the pragmatic aspects of speech acts. As a clarification of my goals, note t h a t the role of the proposed semantics is akin to t h a t of classical semantics for assertives. Classical semantics only tells us when an assertive is objectively satisfied—it makes no claims about when an assertive should actually be uttered or believed. Therefore, the focus here is on satisfaction conditions. T h e conditions of satisfaction for most kinds of speech acts differ significantly f r o m those of assertives t h a t are o r d i n a r i l y considered in logic. Assertives, being claims of fact, are true or false; other speech acts call for a more complex framework in which their felicity or success can be described. In the context of imperatives, Haniblin distinguishes between w h a t he calls exiensional and whole-hearted satisfaction [ H a m b l i n , 1987, p p . 153157]. Briefly, the former n o t i o n admits accidental success, while the latter does n o t . H a m b l i n realized t h a t these were useful things t h a t may be said of a speech act; however, his a i m was simply to be able state prescriptive conditions on when w h a t k i n d of imperatives ought to be issued, and the philosophical problems t h a t arise when one is in a "quandary." T h u s his focus seems to have been p r a g m a t i c . I take advantage of some of his ideas, b u t make a finer distinction and extend it to other i m p o r t a n t kinds of speech acts here, f o r m a l l y rel a t i n g t h e m to intentions and know-how in the process. In §4.2, I address the problem of w h a t k i n d of prescriptive constraints on communication may be stated, b u t see t h a t as essentially supervenient on the semantics. In §2, I discuss three different senses of satisfaction
70
Architectures and Languages
for the five kinds of speech acts considered in this paper. In §3, I describe a theory of intentions and know-how, and a f o r m a l model which I then use to formalize the different notions of satisfaction. In §4, I show how this theory may be used in the design of multiagent systems, and list some useful theorems.
2
Shades of S a t i s f a c t i o n
As remarked above, c o m m u n i c a t i o n among agents in a multiagent system can be best understood by appealing to speech act theory [ A u s t i n , 1962; Searle, 1969]. In speech act theory an " i l l o c u t i o n " (which I identify w i t h a message) is seen to have two parts: an illocutionary force and a proposition. T h e i l l o c u t i o n a r y force distinguishes, e.g., a command f r o m a promise; the proposit i o n describes the state of the w o r l d t h a t is, respectively, commanded or promised. T h i s suggests a simple syntax for messages in an artificial language. A message, m, is a pair ( i , p ) , where i identifies the illocutionary force, and p the proposition. Here i is an atomic symbol f r o m the set {directive, commisive, permissive, proh i b i t i v e , assertive}; and p is a logical f o r m u l a . T h i s much is quite standard even in the AI literature t h a t deals w i t h communication among agents [Huhns et al., 1990; T h o m a s et al., 1990]. However, none of the AI papers so far give a rigorous f o r m a l semantics for messages of different i l l o c u t i o n a r y forces. T h i s lacuna is filled by this paper. B u t before I come to the formalizat i o n , I must discuss the different senses of satisfaction of speech acts. T h e rest of this section extends the discussion in [ H a m b l i n , 1987, pp. 153-157]. Note t h a t these different senses agree for the case of assertives. T h e propositional part of a message specifies the state of the world t h a t the message is, in some sense, about. An assertive asserts of t h a t state t h a t it holds (i.e., currently, though the proposition could be temporally indexed); a directive asks the hearer to b r i n g t h a t state about; a commisive commits the speaker to bringing it about; and so o n . T h u s the satisfaction of a message depends b o t h on its i l l o c u t i o n a r y force and its proposition. T h e different notions of satisfaction are motivated using directives; other speech acts are considered in §3. In the simplest sense of satisfaction, called exiensional satisfaction in [ H a m b l i n , 1987, p. 153], a message is said to be satisfied ( w i t h only m i n o r qualifications) j u s t if its proposition turns out to h o l d . E.g., a directive is satisfied when the proposition becomes true. Extensional satisfaction looks only at the i m m e d i a t e state of the w o r l d — the proposition may have been made true accidentally, or for the " w r o n g reasons," b u t would s t i l l meet the requirements for extensional satisfaction. T h i s notion of satisfaction meets weak behavioral specifications; e.g., if the success of a speech act is p a r t of a p l a n , then when it succeeds the agent can legitimately proceed to the next stage of the plan. However, this n o t i o n is not acceptable for complex systems because f o r t u i t o u s circumstances would n o t , in general, arise often in t h e m . A system whose agents were designed on this basis would t u r n o u t to be not sufficiently robust—e.g., we would expect a guarantee t h a t some directives would be satisfied in a variety of circumstances, rather t h a n t h a t they once
were. Also, it is of not much help practically, since it yields no insights about how the i n d i v i d u a l agents in a multiagent system ought to be designed. T h i s motivates the next sense of satisfaction, wholehearted satisfaction. T h e whole-hearted satisfaction of a directive requires not only t h a t the specified proposit i o n be made t r u e , b u t be made true in a sure-fire manner. T h e concerned agent should not only bring about the r i g h t state of the w o r l d , b u t know how to bring it about and intend to b r i n g it about (thus it would bring it about in a way t h a t exploits its know-how). By using whole-hearted satisfaction, the designer can require t h a t an agent not issue two commands, which cannot b o t h be whole-heartedly satisfied (due to l i m i t e d know-how), even if they can b o t h be extensionally satisfied. Even whole-hearted satisfaction admits cases where the relevant proposition was made true, b u t only because it was going to be made true anyway, irrespective of whether the given speech act was performed. I.e., the speech act was pointless and played no real role in its satisfaction. T h i s happens when, for a directive, the bearer was going to do the desired actions anyway. Often, it is useful to eliminate these conditions, so that the given speech act is really necessary. T h i s requires that not only must the proposition in the speech act be made true in a sure-fire way, it must be made true because of t h a t speech act. T h i s , relevant satisfaction, is the strongest notion of satisfaction t h a t 1 consider here. T h e taxonomy of speech acts of this paper is motivated by the fact t h a t permissives, prohibitives and directives have different satisfaction conditions (cf. [Bach and II arnish, 1979, pp. 39-54] and [Searle and Vanderveken, 1985, ch. 9], where permissives and prohibitives are lumped together w i t h directives). T h e more conventionoriented or culture-oriented illocutionary forces (e.g., christenings, greetings) are not considered here. Interrogatives are semantically quite like directives, b u t need special treatment to allow for answers; they are not i n cluded in this paper for reasons of space.
3
F o r m a l i z a t i o n of S a t i s f a c t i o n Conditions
T h e different notions of satisfaction of speech acts depend on the definitions of know-how and intention. A formal rigorous theory of situated know-how and intentions has already been developed and reported. Here the same framework and technical definitions are used to give an account of the different sorts of satisfaction of several kinds of speech acts. T h i s project has been inspired by t h a t of C. L. H a m b l i n , who died while w r i t ing [ H a m b l i n , 1987]. He argued t h a t an account of i m peratives must be b u i l t on top of a theory of abilities and intentions, especially, one which is " . . . n o t hidden in the m i n d , however, b u t expressed in action . . . " [Hamb l i n , 1987, foreword by Belnap, p. v i i i ] . He d i d not have such a f o r m a l theory of know-how and intentions, b u t I take this advice seriously, and t r y to give a semantics of several kinds of speech acts in terms of my earlier theory of intentions and know-how, where these concepts are defined in terms of the actions of agents situ-
ated in an objective model [Singh, 1990b; Singh, 1990a; Singh, 1991]. I briefly describe that theory, and then t u r n to the formal model. 3.1
Know-how and Intentions
T h i s theory of know-how and intentions is meant to apply to b o t h t r a d i t i o n a l plan-based architectures and modern situated ones. It uses the concept of strategies as abstract descriptions of the agents' behavior. Strategies correspond to plans in t r a d i t i o n a l systems and to the architectural structure of reactive agents, as instantiated at a given t i m e . A strategy is simply the designer's description of the agent and the way in which it behaves. An agent knows how to achieve A, if it can achieve A whenever it so "intends." Let each agent have a strategy t h a t it follows in the current situation. Intuitively, an agent knows how to achieve A relative to a strategy Y, iff it possesses the skills required to follow Y in such a way as to achieve A. Thus know-how is partitioned into two components: the " a b i l i t y " to have satisfactory strategies, and the " a b i l i t y " to follow them. Similarly, an agent intends to achieve A by performing a strategy Y, iff it currently has strategy Y and the successful performance of Y by it entails A; i.e., iff the agent is t r y i n g to perform Y to achieve A. Note t h a t having an intention does not entail having the know-how to match i t , and vice versa. Another i m p o r t a n t p r i m i t i v e is can-prevent, notated Kprev. T h i s is related to know-how and applies when the given agent is able to perform actions so as to prevent the occurrence of the given condition. For reasons of space, the technical details of [Singh, 1991; Singh, 1990a; Singh, 1990b] are not included here. The presentation below is self-contained, however. 3.2
The Formal Model
T h e formal model here is based on possible worlds. Each possible world has a branching history of times. Agents influence the future by acting, but the outcome also depends on other events. A scenario at a given world and time is any linear branch of the f u t u r e beginning there— this corresponds to a particular run or trace of the given system. A subscenario is a t r i p l e , ( S , t , t ' ) , which denotes a section of scenario S f r o m t i m e t to t'. T h e interpretation [] assigns sets of world-times pairs to predicates, for each possible tuple of their arguments; it assigns sets of subscenarios to each action, for each agent who might do i t . T r u t h in the model, M, is defined relative to a world w and a t i m e t: M \=w,t P denotes t h a t p is true in M at w and t. Another useful notion is of t r u t h relative to a scenario and a t i m e in it (a scenario determines the world): M \=s,t P denotes t h a t p is true in M on S at t. T h e formal language here is the predicate calculus, augmented w i t h temporal logic (used in §4.3), and three predicates, 'K h o w , 'intends' and 'K'pre,,,' each applying to an agent, a strategy and a f o r m u l a . Speech acts are, first of a l l , actions. For simplicity, they are seen as the actions of j u s t their speakers, and occur over subscenarios. Let 'says-to' be a parametrized speech act, to be used as 'says-to(y, m).' T h i s action w i l l be seen as an action done by agent x. (ts,,th) E [ s a y s - t o ( y , m ) ] x means t h a t agent x commu-
Singh
71
nicated message m to agent y in the t i m e f r o m t s (the time of speaking) to t h (the t i m e of hearing). T h i s j u s t says t h a t the message was successfully t r a n s m i t t e d ( i t is possible to allow failed transmissions, b u t t h a t is not useful here). Define a new predicate ' c o m m ' t h a t applies to two agents, and a message. ' C o m m ( x , y , m ) ' is true at w, t j u s t if x said (or started to say) rn to y then. A t r a n s m i t t e d message may, of course, not be satisfiable. In order to be able to t a l k of the different kinds of satisfaction of messages, I introduce three operators: ESAT, WSAT and RSAT (collectively called SAT below) t h a t apply on formulas of the f o r m ' c o m m ( i , y , m ) . ' They respectively state t h a t the given message is extensionally, whole-heartedly, and relevantly satisfied. Now the conditions of t r u t h are given for each of these three operators and for each of the possible illocutionary forces, relative to a scenario and a t i m e .
3.3
Extensional Satisfaction
Extensional satisfaction is defined relative to a scenario and a time in it (thus " f u t u r e " means future w i t h i n t h a t scenario). A directive is satisfied at a scenario and time j u s t if its proposition becomes true at a future time on t h a t scenario. A commisive too is satisfied j u s t if its proposition becomes true at some future t i m e on the given scenario. An assertive is satisfied j u s t at those times where its proposition is true. A permissive is satisfied on a scenario and t i m e j u s t if it is taken advantage of sometimes in the f u t u r e of t h a t t i m e on t h a t scenario. A p r o h i b i t i v e is satisfied j u s t if it is never violated in the f u t u r e of the given t i m e on the given scenario. T h u s the agents' intentions and know-how do not m a t t e r for extensional satisfaction.
3.4
Whole-hearted Satisfaction
Whole-hearted satisfaction too is defined relative to a scenario and a t i m e . A directive is satisfied on a scenario
72
and t i m e j u s t if its proposition becomes true at a future t i m e in t h a t scenario, and all along the scenario f r o m the given t i m e to t h e n , the hearer has the know-how, as well as the i n t e n t i o n to achieve i t . I.e., if the hearer has a strategy ( t h a t it may be said to be following) relative to which it has the know-how and the intention to achieve p (as explained in §3.1). Similarly, a commisive is satisfied j u s t when its proposition becomes true at some f u t u r e t i m e on the given scenario, and all along the scenario f r o m the given t i m e to t h e n , the speaker has the know-how to achieve it and also intends i t . T h e condition for assertives is unchanged. A permissive is satisfied at a scenario and a t i m e j u s t if it is taken advantage of by the hearer at a f u t u r e point on t h a t scenario. B u t when a permissive is taken advantage of, it allows the hearer to do actions at certain times t h a t it could not have done before because they m i g h t possibly lead to the c o n d i t i o n becoming true. Thus a permissive is satisfied on a scenario on which the hearer does at least one action whose performance can lead to a situation where it is unable to prevent t h a t condition f r o m occurring (i.e., the hearer can now risk l e t t i n g t h a t condition hold). Similarly, a p r o h i b i t i v e is satisfied at a scenario and t i m e j u s t if none of the actions done by the hearer on t h a t scenario ( i n the f u t u r e ) , can lead to a situation where the hearer would be unable to prevent the condition f r o m occurring (i.e., the hearer does not risk v i o l a t i n g the p r o h i b i t i o n ) .
3.5
Relevant Satisfaction
Relevant satisfaction is also defined relative to a scenario and a t i m e . It resembles the previous case, b u t differs in adding a requirement t h a t roughly says t h a t the given speech act is the true reason for its success. A directive is satisfied j u s t when its proposition becomes true at a t i m e in the f u t u r e of the given t i m e , and all along the scenario f r o m the given t i m e to t h e n , the hearer has the know-how to achieve i t , and furthermore t h a t this know-how does not arise in at least one scenario t h a t is a t e m p o r a l alternative to the given one. T h e definition for commisives is analogous. T h e condition for assertives continues to be the same as before. A permissive is satisfied at a scenario and a t i m e j u s t if it is taken advantage of by the hearer at a f u t u r e t i m e on the given scenario, w i t h an a d d i t i o n a l requirement t h a t
4
Applying the Theory
T h e two m a i n m o t i v a t i o n s for developing this theory are to provide a rigorous f o u n d a t i o n for the design of m u l tiagent systems and to j u s t i f y some prescriptive claims about how agents should communicate in t h e m . T h i s theory has given objective criteria to evaluate the correctness of different scenarios, or runs of a multiagent system. In design, the problem is to create a system which allows only correct scenarios to be actualized. Prescriptive claims for agents tell them what to do given their beliefs and intentions, so t h a t correct scenarios emerge. I discuss these two problems below, and then some i m p o r t a n t f o r m a l consequences of this theory. 4.1
Designing M u l t i a g e n t Systems
have (at the appropriate times) (1) the know-how to answer correctly whether they w i l l b i d on a given task, and to promise as described above; and (2) the corresponding intentions. Thus the design goal reduces to ensuring conditions (1) and (2) for all the agents. T h e different senses of SAT yield rigorous definitions for three kinds of correctness conditions for multiagent systems. A scenario can be defined to be correct in the sense of SAT if all the messages passed on it are satisfied in the same sense. In general terms, the designer's goal is to ensure t h a t all runs t h a t may be actualized are correct. T h i s reduces to the goal t h a t the intentions and know-hows of the agents are such t h a t only correct scenarios are actualized. T h i s is the sense of correctness t h a t designers use in practice. They usually achieve this k i n d of correctness by, e.g., h a r d - w i r i n g the intention to cooperate in their agents, or by setting up hierarchical structures such t h a t some directives (commands) are always obeyed, and others (requests) obeyed whenever they do not conflict w i t h the hearer's current intentions. 4.2
Normative Constraints on Communication
T h e objective criteria given above can be used to motivate some normative constraints on communication among the agents in a multiagent system. These constraints could be used by the designer, and possibly by the agents themselves to reason about the messages they exchange. Note t h a t these are j u s t meant as weak constraints, and may be easily overridden. If (some of) these constraints are obeyed, the scenarios t h a t are actualized are not j u s t correct b u t also " g o o d . " Imposing these constraints can simplify a design since certain good properties can then be taken for granted. 1. An agent should issue a directive only if its intentions are satisfied (i.e., its current strategies arc performed successfully) on all scenarios on which the directive is W S A T . T h i s is the converse of the defi n i t i o n for M )=w,t WSATq suggested in §3.2—this ensures t h a t actions by the hearer w i l l not render the speaker's own intentions impossible to achieve.
One extension needed is to distinguish between messages of the same m a j o r illocutionary force t h a t are different in some i m p o r t a n t respects. E.g., commands differ f r o m requests, since they presuppose a u t h o r i t y on part of the speaker, and can cause a change in the hearer's intentions under a wider variety of circumstances t h a n requests. A correct scenario w i t h a command must WSAT it as a directive (assuming a u t h o r i t y ) ; one w i t h a request must WSAT it t o o , b u t only if some other conditions on cooperation are met. T h e designer needs to constrain the issuing of directives, a n d / o r increase know-how and constrain intentions so t h a t the system actualizes only correct scenarios. E.g., requiring t h a t agents persist w i t h their strategies for sufficiently long considerably s i m p l i fies generating runs t h a t on which directives and commisives are W S A T . Correctness must be ensured for all the message types t h a t can occur in the system.
3. Agents ought to persist w i t h their strategies (i.e., their strategy at a later t i m e should be the appropriate " t a i l " of their strategy at an earlier t i m e ) . T h i s not only simplifies the W S A T and RSAT conditions for directives and commisives, b u t also s i m p l i fies the interactions among agents by, e.g., m a k i n g it easy for an issuer of a directive to not take on a strategy t h a t would interfere w i t h its compliance.
As an example, consider the contract net [Davis and S m i t h , 1983]. In its simplest f o r m , a manager sends out a call for bids to all contractors—treat this as requesting the hearer to b i d and to promise t h a t it w i l l do the task for a certain price ( i f it wants to b i d ) . Contractors who b i d thus promise t h a t if requested to do the task and given their price, they w i l l do i t . T h e manager selects one contractor and requests it to do the task. In any correct design for this protocol, the contractors must
4. A l l messages sent by one speaker must be m u t u ally consistent in the sense of being j o i n t l y satisfiable on at least some scenarios. E.g., different d i rectives should not clash w i t h each other and prohibitives should not preclude the satisfaction of d i rectives and other prohibitives. T h i s prevents many unacceptable situations, b u t can cause problems if some redundant permissions are issued, i.e., those t h a t m i g h t never be used on some scenarios where
2. A l l messages sent must be RSAT, where the quantification over scenarios is restricted to scenarios compatible w i t h the speaker's current intentions.
Singh
73
References [Allen and Perrault, 1980] J. F. A l l e n and C. R. Perrault. A n a l y z i n g intention in utterances. Artificial Intelligence, 15:143-178, 1980. [ A u s t i n , 1962] J. L. A u s t i n . How to do Things with Words. Clarendon, O x f o r d , U K , 1962. [Bach and Harnish, 1979] K. Bach and R. M. H arnisn. Linguistic Communication and Speech Acts. MIT Press, Cambridge, M A , 1979. [Cohen and Levesque, 1988] P. R. Cohen and H. J. Levesque. R a t i o n a l interaction as the basis for comm u n i c a t i o n . Tech. Rep. 433, S R I I n t e r n a t i o n a l , Menlo Park, C A , A p r . 1988. [Davis and S m i t h , 1983] R. Davis and R. G. S m i t h . Neg o t i a t i o n as a metaphor for distributed problem solving. Artificial Intelligence, 20:63-109, 1983. [Emerson, 1989] E. A. Emerson. Temporal and modal logic. In J. van Leeuwen, ed., Handbook of Theoretical Computer Science. N o r t h - H o l l a n d Publishing Company, A m s t e r d a m , 1989. [Gasser and Huhns, 1989] L. Gasser and M. N. Huhns, eds. Distributed Artificial Intelligence, Volume II. P i t man, London,1989. [Grice, 1969] P. Grice. Utterer's meaning and intentions. Philosophical Review, 1969. [ H a m b l i n , 1987] C. L. H a m b l i n . Imperatives. Blackwell L t d . , O x f o r d , U K , 1987.
Basil
[Huhns, 1987] M. N. Huhns, ed. Distributed Artificial Intelligence. P i t m a n , L o n d o n , 1987.
5
Conclusions
None of the extant theories of communication address quite the same problem as I have addressed in this paper. T h e theory presented here refines and formalizes some i n t u i t i o n s about communication among agents. 1 started o u t w i t h a set of obvious and well-known i n t u itions about the nature of communication derived f r o m classical speech act theory. However, I motivated and approached an i m p o r t a n t problem t h a t has not been addressed in the literature on speech act theory, or even in the AI literature. My c o m m i t m e n t is stronger to the general claims t h a n to the specifics. However, using definitions of the intentions and know-how of an agent, I was systematically able to give rigorous definitions of the conditions of satisfaction for speech acts of different of i l l o c u t i o n a r y forces. These definitions capture many of our i n t u i t i o n s about when, as speakers and hearers, we believe t h a t a given speech act has been satisfied. T h i s theory has applications in the design of multiagent systems, where constraints on the know-how and intentions of agents are derived f r o m their desired communicative behavior. It can also yield some well-motivated normative constraints on communication among agents. An advantage of the f o r m a l approach is t h a t this process can be guided by the several theorems t h a t exist.
74
Architectures and Languages
[Huhns et al., 1990] M. N. Huhns, D. Bridgeland, and N. A m i . A D A I communication aide. Tech. Rep. A C T RA-317-90, Microelectronics and Computer Technology C o r p . , A u s t i n , T X , Oct. 1990. [Perrault, 1987] R. Perrault. An application of default logic to speech act theory. Tech. Rep. 90, C S L I , Stanf o r d , C A , Mar. 1987. [Searle and Vanderveken, 1985] J. R. Searle and D. Vanderveken. Foundations of Illocutionary Logic. Cambridge University Press, Cambridge, U K , 1985. [Searle, 1969] J. R. Searle. Speech Acts. Cambridge U n i versity Press, Cambridge, U K , 1969. [Searle, 1975] J. R. Searle. Indirect speech acts. In P. Cole and J. L. M o r g a n , eds., Syntax and Semantics, Volume 8. Academic Press, New Y o r k , N Y , 1975. [Singh, 1990a] M. P. Singh. G r o u p intentions. In 10th Workshop on Distributed AI, Oct. 1990. [Singh, 1990b] M. P. Singh. Towards a theory of situated know-how. In ECAI, A u g . 1990. [Singh, 1991] M. P. Singh. A logic of situated know-how. In AAAI, J u l y 1991. [Thomas et al., 1990] B. T h o m a s , Y. Shoham, and A. Schwartz. Modalities in agent-oriented p r o g r a m m i n g . C o m p u t e r Science D e p a r t m e n t , Stanford University, Feb. 1990.
