Published in Jennings, N. and O'Hare, G. (eds.), Foundations of Distributed Artificial Intelligence , Wiley, New York (1996), pp. 487-503

(Note: This is a manuscript version. Please quote only the published version.)


PHILOSOPHY AND DISTRIBUTED ARTIFICIAL INTELLIGENCE
The Case of Joint Intention

Raimo Tuomela
University of Helsinki


I PHILOSOPHY AND DAI

In current philosophical research the term 'philosophy of social action' can be used - and has been used - in a broad sense to encompass the following central research topics: 1) action occurring in a social context; this includes multi-agent action; 2) joint attitudes (or "we-attitudes" such as joint intention, mutual belief) and other social attitudes needed for the explication and explanation of social action; 3) social macro-notions, such as actions performed by social groups and properties of social groups such as their goals and beliefs; 4) social norms and social institutions (see Tuomela, 1984, 1995). The theory of social action understood analogously in a broad sense would then involve not only philosophical but all other relevant theorizing about social action. Thus, in this sense, such fields of Artificial Intelligence (AI) as Distributed AI (DAI) and the theory of Multi-Agent Systems (MAS) fall within the scope of the theory of social action. DAI studies the social side of computer systems and includes various well-known areas ranging from Human Computer Interaction, Computer-Supported Cooperative Work, Organizational Processing, Distributed Problem Solving to Simulation of Social Systems and Organizations. Even if I am a philosopher with low artificial intelligence I will below try to say something about what the scope of DAI should be taken to be on conceptual and philosophical grounds. (In the later sections of the paper the central notion of joint intention will be the main topic - in order to illustrate how philosophers and DAI-researchers approach this issue.)
Let us now consider the relationship between philosophy - especially philosophy of social action - and DAI. Both are concerned with social matters and in this sense seem to have a connection to social science proper. What kinds of questions should these areas of study be concerned with? In principle, ordinary social science should study all aspects of social life (in various societies and cultures), try to describe it and create general theories to explain it. Philosophers do not create specific social theories but their task in this context is to create conceptual frameworks for the study of social life, to investigate central social concepts, and to discuss and critically evaluate different fundamental ideas and philosophical views about social life and social systems. What is the task of DAI, then? Broadly put, DAI should basically attempt to study - and to create "design specifications" for - a social world of intelligent robots; it should also be concerned with the actual technical implementation of those specifications. As DAI is basically concerned with intelligent agents it typically adopts rationality assumptions which are stronger than those social scientists can plausibly make (social scientists should also be interested in stupid, clumsy and mistake-making agents), but because of the end of actually building robots it will have to take into account the resource-boundedness of intelligent agents. This concern with bounded resources may also contrast with how philosophers - those prime idealizers - approach these matters. My most general thesis in this paper is this: Philosophy is in many ways highly relevant and important for DAI, and - to some extent - vice versa . A sharp borderline between these fields hardly exists, even if philosophers of social action, qua philosophers, are interested in somewhat different aspects of social action (and related matters). I argue that, in principle, one of the tasks of philosophy is to specify the research areas and problems of DAI on a general level (but only on a general level). DAI need not be interested in simulating all the inner mechanisms and processes that lead agents to do what they do in a social setting, but it has to be concerned with what people achieve by their actions - with "social outputs" - and various related public matters including social reasons for action. This public output aspect is something that philosophers should specify on a general level: Philosophers ought to say what the output side of social life - and the life of the community of intelligent robots - basically is, or at least they ought to be able to sketch plausible scenarios about it. Here we arrive at fundamental questions such as what general reasons there are for people to form communities and societies at all and what kind of conceptual problems are involved in answers to these questions. This is a broad topic which I can only touch briefly here.
Basically, agents have not only personal needs and wants but also collective ones. Thus we may have as our collective goal to keep our environment clean; no single agent alone can achieve that collective goal. Or, to mention a small-scale collective or joint goal, we may want to get deer meat, believing that cooperative joint hunting is desirable or even necessary as compared with hunting separately or alone. In this kind of context - especially when a collective goal or "good" is to be achieved - collective action problems are central. Such problems involve a conflict between individual and collective interests, because there is an inbuilt incentive to free-ride (e.g. in our example to litter or not to care). Collective action problems typically have the structure of a Prisoner's Dilemma game or a Chicken game.
As indicated, in social life there are central tasks that can be solved only collectively: a) tasks requiring collective goals (e.g. the provision of collective goods in the economist's sense); often but not always there is a collective action problem here - conflict between collective and individual rationality; b) individual goals (tasks) which, however, require collective cooperation concerning the means (personal goals thus can be satisfied only through collective or joint action). Collective needs and collective goals over and above personal ones thus are central for the formation and maintenance of societies and social institutions, and in this context collective action problems typically occur (norm-based social institutions are supposed to help to solve them). There is of course much more to social life, but basically the above considerations, when spelled out, should cover the most central aspects. As a result we get a conceptual framework of public, output-related social concepts which will include also notions related to social positions, social norms, and social power over and above explicit action notions.
Roughly in this way we then get the needed "output specifications" (understood in the above broad sense), and then also the scope of DAI on a general level becomes determined. As to the question of how to theorize about these matters and how ultimately to create a theory satisfying those output specifications, space does not allow me to say much here. Let me note, however, that here also philosophy has a role to play: it can be assigned the task to create a broad conceptual framework of concepts and basic underlying assumptions for the study of central aspects of social life. This task - because of its generality and broad scope - goes beyond what social theorists can be expected to do. What kinds of elements does such a conceptual framework contain? Here is a short and incomplete list of notions that philosophers have studied: joint intention, mutual knowledge, joint and collective action, joint decision-making, social norms of various kinds, social institutions, agreement and contract, communication, bargaining, power, strategic interaction, love and caring, morality, social values and good society (many of these concepts have been investigated in Gilbert, 1989, Pettit, 1993, and Tuomela, 1995). These notions - central for the study of social life - will have to be ultimately taken into account also by DAI.
I have claimed thus that philosophers ought to create relevant social conceptual frameworks and show DAI in broad terms where to go, what kind of problems to undertake to solve or at least indicate in what kind of conceptual setting or settings feasibly to work. (It is of course clearly the task of DAI rather than philosophy to determine what its specific research problems ought to be - that depends on the internal development in DAI and in its neighbor sciences.) Let me idealize the situation somewhat and make the following slightly speculative suggestion in terms of the software-hardware distinction: Think of the (or a) public social framework as a kind of software in terms of which to theorize, while the hardware will consist of the agents "realizing" that software and "computing" the programs stated in terms of that software. The present software-hardware distinction is not quite clear-cut: While idiosynchratic psychological states and processes clearly would belong to the hardware side possibly also some non-general psychological information should be put on the social software side in a publicly codified form (e.g. in the form psychological types - altruist, free-rider, dominant).
To continue with the tasks of philosophy, it is also expected to sketch the viable philosophical options (or at least arguemnts pro et con ) and investigate their features. For instance, in the case of the mind-body problem it will specify the various versions or monism, materialism (or physicalism) and dualism and discuss their features. (Examples of some recently much discussed mind-body views are identity theory, token physicalism, supervenience physicalism, and functionalism.) Each DAI-researcher is committed to, and relies on, some philosophical background assumptions - whether he is aware of it or not. The task of philosophy is also to show how DAI connects to other sciences and disciplines such as the brain sciences.
Given the above remarks it should be clear that the social conceptual framework that DAI will ultimately need is quite broad. Thus, for instance, the Belief-Desire-Intention (BDI) architecture that has been employed in the study of both single-agent and multi-agent action will have to be complemented by other social concepts, as indicated. Over and above helping in the creation of conceptual frameworks and in the critical investigation of underlying assumptions and arguments philosophy can also help DAI in more down-to-earth questions related to actual everyday research. Thus philosophy can contribute to DAI in the research problems related to logic, in the analysis of specific notions and topics, questions of rationality, and in making choices concerning metaphysical options (e.g. formal ontology, limits of DAI). Analytic philosophers are also good at inventing substantive, unartificial examples - especially counterexamples to specific theses.
As researchers in AI know, they owe much of their logical and exact tools to philosophy. Especially, modal logic, which is so important in AI today, was first developed by philosophers. Possibly most of the research in modal logic is now conducted by researchers within AI, but also philosophers continue to work in that field. I will not present evidence for this - one only needs to open a textbook in AI to see the relevance of the work by philosopher-logicians.
Recent developments within the philosophy of mind and the philosophy of action have helped the concept formation in AI and in part vice versa (cf. the frequent references to books such as Bratman, 1987 and the work on belief-kinematics, and - in the other direction -the influence of AI on the philosophy of mind and even philosophy of science). As for DAI and multi-agent theory, one can find traces of influence - and references - in DAI literature to philosophers' writings, although it must be said that even within philosophy the questions related to the theory of social action were not seriously researched before the 1980's. Some relevant pieces of philosophical work related to joint intention and joint action, cooperation, and communication that also some DAI-researchers have noticed are the papers by Bratman (1992, 1993), Grice (1957), Searle (1990), and Tuomela (1991, 1993a) as well as Tuomela and Miller (1988). Other relevant pieces of logical and philosophical work are the works by Belnap (1991), Bratman (1994), Cohen and Levesque (1991), Gilbert (1989), Tuomela and Sandu (1994), Sandu and Tuomela (1994), Tuomela (1984, 1993b, 1995). (Relevant works by DAI-researchers will be referred to and discussed later in this paper.) As to questions of rationality and metaphysics - mentioned above - there has not yet been much interaction between philosophy and DAI. It is easy to predict, however, that questions of individual and collective rationality - and rationalization - will have to be tackled also by DAI-researchers in the near future when the study of robot communities really gets on its way. Then also all kinds of questions related to the metaphysics of the social realm will have to be confronted (e.g., the reality of groups and social structures).


II JOINT INTENTION: THE ANALYSANDUM AND CONDITIONS FOR ANALYSIS

To show that there is philosophical research very close to research in DAI, I will in the rest of this paper discuss the central multi-agent notion of joint intention (and to a lesser extent that of joint action). The material I will present is rather concise, and in many places I cannot use space for presenting arguments but refer the reader to other writings. While I will concentrate on my own research below, some representative examples of other theoreticians' work will also be considered towards the end of the paper. The purpose of the discussion in the rest of the paper is nevertheless not to give a comprehensive survey. Rather it is to show in a detailed fashion how philosophers and DAI researchers can work hand in hand in this kind of context.
The theoretical view I will present to you has as its intuitive analysandum the following situation. Suppose one of us comes up with the idea of cleaning our backyard. He may express this idea colloquially by saying "Let's clean our backyard! Are you with me?" The others may answer "Yes" or "A good idea!". Suppose there are three persons involved all of whom in this sense accept the idea of cleaning the backyard. We can say that what they accept is a plan for doing something together. Another way of putting this matter is that they have made an agreement to clean the backyard. The intuitive analysandum here is, anyway, that of accepting - separately or, in some cases, jointly - a plan to do something together (cf. "Here is a plan for joint action. Who accepts it"?).
In my analytic terminology we can say the following about the analysandum. The participants of a joint action to be intentionally performed have publicly accepted a plan for performing that action (or at least a plan somehow implying the performance of that action). The joint action thus is plan-based, and the participants' having and endorsing that plan involves their joint intention to perform the action in question. In my analysis I will accordingly say that the participants conatively endorse a plan involving or entailing the joint-intention expression "We will do X together", X being a joint action. (When applied to a certain participant I will say that he has or shares the "group-intention" to do X; what will be called a "we-intention" is a special kind of group-intention.) The endorsement of the plan must be public and in standard cases it must also be communicated to the other participants. The acceptance of the plan by the participants can take place purely individually - cf. "This is the plan. Who accepts it?". The acceptance of the plan, however, commits the participating persons to action - this is commitment in a strong interpersonal sense concerned both with the joint action and the participants' parts or shares, with responsibility to take part towards the other participants.
I have elsewhere (in Tuomela, 1995) argued that joint intentions are central in several respects. As I will below be mainly concerned with analysing the notion of joint intention, it is useful to present a list of features that indicate the centrality of joint intentions and also serve as a kind of criteria of adequacy that any analysis of joint intention should respect. Consider thus the following list (also cf. Bratman, 1992):
1) Joint intentions figure centrally in participating agents' (social) practical reasoning and resulting joint action. Thus:
a) Joint intentions - in analogy with single-agent intentions - present problems and restrict available action alternatives. Forming a joint intention to go swimming excludes other alternatives and present problems concerning the means to perform the joint action in question.
b) Joint intentions serve to initiate, guide, and monitor joint action, creating order - especially interpersonal coordination - in social life. Furthermore, there cannot be intentional joint action without joint intention (at least joint "action-intention").
c) Joint intentions help to connect the group-level with the personal level. Thus it can be argued that a group's intentional action - e.g. a business company's buying another company - must be based on the joint intentions of some authority-possessing members of the group.
d) Joint intentions have a normative impact on the agents' thinking and acting. (Recall my remarks on joint commitment.)
2) The concept of joint intention is central both in philosophy of social science and in theorizing about the social world (and thus it occurs in some sense in various theorists' heads, so to speak). Thus:
a) The concept of joint intention (a person's we-intention) is needed for a characterization of the social "I" (viz. "I" as "one of us"). It will argued below that joint intentions - which are necessary ingredients of intentional joint action - are irreducible to personal intentions and other personal notions.
b) The concept of joint intention helps to conceptualize the conflict between the group-level and personal level interests - cf. collective action dilemmas - and is central for the analysis of group-phenomena such as actions performed by groups in terms of what group members think and do;
c) The notion of joint intention (and related jointness-notions) can be and should be important in theory-formation in the social sciences. The most basic argument for this is simply that people have joint intentions in many central social situations and that therefore social theory - and also DAI - must deal with the concept of joint intention.
d) The notion of a social group, in a refined sense, is basically an "authority-involving" system for arriving at joint intentions involving group commitments (see Tuomela, 1995, for discussion).

As said, the above features also qualify as a kind of criteria of adequacy for any account of joint intention. (That criteria 1)a)-d) will be fulfilled by the account to be offered should be rather clear from my presentation in the next section; as for 2)a)-d) I refer the reader to Tuomela, 1995.)


III AN ANALYSIS OF JOINT INTENTION

Suppose that you and I have the joint intention to clean our backyard together, as expressed by the conative sentence "We will clean our backyard together". The content of the intention here involves our performing something together and the pronoun 'we' of course refers to our group. A joint intention (expressible by a sentence of the form "We will do X"") is something that several agents have among themselves.
In contrast to wants, wishes, and hopes, and many other mental states, intentions are, in the last analysis, necessarily related to one's own action. Accordingly, when an agent is said to take part in a joint intention, what he ultimately will have to intend is that he by his own action, by his part or share, contribute to the joint action. A fully intentional joint action X (cf. jointly carrying a table, singing a song) must arguably be based on the participants' relevant joint intention (normally the intention to do X), and we should accordingly require that each participant have the intention to do his part of X. This kind of intention is just an intention that can be called a we-intention.
Intentions being in part cognitive, an agent cannot intend to perform an action unless he believes that this action is possible or at least realizable with some nonnegligible probability. With double analogy, an agent cannot we-intend unless he believes not only that he can perform his part of X but also that he together with his fellow participants can perform the joint action in question: The jointly intending agents must believe that the "joint action opportunities" for an intentional performance of X are (or will be) there. Yet another property of a we-intention is that in each participant's view it must be mutually believed by the participants that the presuppositions for an intentional performance of X hold. Even more must be said, if my plan-view of intention is right: The agents must believe that they have accepted a plan for joint action, and at least in standard cases this must be just the we-intended joint action. What this involves is that the concept of we-intention is conceptually dependent on a preanalytic notion of joint intention. Thus we-intentions cannot be analyzed in terms of personal intentions and mutual beliefs. It can be debated whether the dependence of we-intention on joint intention is of the nature of conceptual presupposition or whether it should be built into the concept of we-intention. If we accept the latter view we must accept that it is a conceptual truth in the case of a we-intending agent A that he we-intends to do X in accordance with and because of the participants' joint intention to do X. This can be argued for by reference to the satisfaction conditions of the we-intention: Unless A performed his part of X because of the participants' plan to do X he surely would not have satisfied his we-intention.
Actually joint intentions must be taken to include also standing intentions. We-intentions directly generate action and do not include all standing joint intentions, e.g. those concerned with future actions about whose performance ("joint action opportunities") there is no definite belief yet. Group-intentions in this sense can be regarded as precisely those intentions which are expressible by locutions of the form "We will do X" (where 'will' is conatively used). Arguably, such group-intentions are either we-intentions or dispositions to we-intend, and I have argued that joint intentions can be regarded basically as mutually shared we-intentions (or at least dispositions to we-intend) about whose existence the participants have mutual knowledge (or belief). (See Tuomela, 1991, and especially 1995, Chapter 3.)
In contrast to joint desires and wants, we can say the following about joint intentions (shared we-intentions): Based on their various personal and, especially, joint desires and mutual and other beliefs, the agents make up their minds - jointly or separately - and thereby form an intention (or plan, as we may say) to act jointly, attempting to act rationally in a coordinated way so as to fulfil the already formed plan. Joint intentions, in contrast to joint wants and desires, can be regarded as joint commitments to action, viz. the participants' interdependent commitments to perform their parts of the joint action and their responsibility for the total joint action getting performed. (We can also say more generally that the participants are jointly committed to reach a joint goal - such as X's having been performed jointly by them - and their plan to achieve this joint goal, as reflected in their part-performances.) Such joint commitments are appropriately persistent and, especially, are not consummated before the agents have jointly achieved what they we-intend (or achieved consensus about the unachievability of the intended goal). However, as in the case of mere personal intentions, the agents can change their mind, and thus joint commitments are not irrevocable. It is nevertheless a rationality condition that the participants preserve their we-intentions as long as the others do or else at least inform the others about their changes of mind in order to become released from their commitment to participate in the intended joint action.
Let us now think of some agents' having accepted "We will jointly clean our backyard" and thus agreed to act jointly. This acceptance must be conative - they must be disposed to contribute to the joint task and also be normatively committed to that task. Joint intentions are formed in such a process of agreement-making (not all agreement-making needs to result in a joint intention). In joint-intention formation each agent also accepts for himself: "I ought to participate in our doing X together". This acceptance means not only that the agent (at least dimly) recognizes the existence of the agreement ("plan") to perform X but also commits himself to performing X together with the others.
The agreement-making that takes place in the context of the formation of a joint intention can take place fully explicitly or only implicitly. Agreement-making - be it explicit or not - is an essentially intentional activity and also a communication-based public activity. Mistakes and errors about the contents of agreements can of course occur, but in the standard cases the participants will share the correct mutual belief not only about that they have made an agreement but also about its content. According to the agreement-view of joint intention and joint action every full-blown or "proper" intentional joint action and - accordingly joint intention - must be based on an agreement by the participants. However, this is only the core idea. There are cases of a) pre-agreement and b) institutional cases in which no agreement-making concerning the particular action in question is needed. (I shall not here discuss these issues further - see Tuomela, 1995, Chapter 2.)
Let us consider the central notion of we-intention in some more detail. We-intentions are action-generating joint intentions that agents have in situations of joint action, e.g., when they intend to carry a table jointly. The content of a we-intention can (with some reservations) be taken to be "to do X jointly" (or, to emphasize its "we-character", something like "we to do X jointly"). A we-intention involves the intention to perform one's part of the joint action. I will now present a detailed summary analysis of the notion of we-intention (cf. Tuomela, 1995). We consider a member Ai of a collective G ("we" for Ai) assumed to consist of A1,...,Am, with i =1,...,m (and allowing the word 'agreement' to refer to merely believed agreement):

(WI ) A member Ai of a collective G we-intends to do X if and only if, based on the (explicit or implicit) agreement to perform X jointly made by the agents A1,...,Ai,...,Am,
(i) Ai intends to do his (agreement-based) part of X (as his part of X);
(ii) Ai has a belief to the effect that the joint action opportunities for an intentional performance of X will obtain (or at least probably will obtain), especially that a right number of the full-fledged and adequately informed members of G, as required for the performance of X, will (or at least probably will) do their parts of X, which will under normal conditions result in an intentional joint performance of X by the participants;
(iii) Ai believes that there is (or will be) a mutual belief among the participating members of G (or at least among those participants who do their parts of X intentionally as their parts of X there is or will be a mutual belief) to the effect that the joint action opportunities for an intentional performance of X will obtain (or at least probably will obtain);
(iv) (i) in part because of (ii) and (iii).

In this analysis clause (i) is the most central one; (ii) and (iii) can be regarded as presupposition-beliefs, whose effectiveness is guaranteed by (iv).
We actually need a general notion of joint intention (or group-intention), viz. a notion which covers not only action-generating but also standing joint intentions. Put somewhat differently, we need a characterization of the notion of joint intention satisfying the truth conditions of the conatively used, intention-expressing sentence "We will do X", where X denotes an arbitrary joint action type. While it is rather obvious that we-intentions in the sense of our (WI ) make this sentence schema true, there are also other truthmakers. I have argued in Tuomela (1991) that conditional we-intentions with conditions believed by the agent to be satisfied also qualify, and so also do dispositions to we-intend in general. More exactly, I defend essentially the following analysis:

(GI ) "We will do X" is true of A (relative to A's group G) if and only if, based on the (explicit or implicit) agreement by the members of A's group G to perform X jointly,
1) A we-intends to do X (in the sense of (WI )); or
2) A has formed a standing group-intention to do X, which is a disposition to we-intend to do X (provided X has a part-division).
Group-intentions can now be regarded as either we-intentions or dispositions to we-intend. On the basis of our earlier discussion we can now give our analysis of the notion of joint intention:

(JI ) Agents A1,...,Ai,...,Am have the joint intention to perform a joint action X if and only if
a) these agents have the group-intention to perform X; and
b) there is a mutual belief among them to the effect that a).
The upshot of our present analysis is this. Joint intentions are intentions that several agents among themselves have, and they are expressible by an intention-expression of the form "We will do X" endorsed by these agents. When two or more agents have the joint intention to do X each of these agents also accepts the intention-expression "We will do X" and essentially because of that they can be said to have the group-intention to do X. Furthermore, when the joint action X has a - preassigned or presently agreed-on - part-structure, the an agent's group-intention to do X is either his we-intention to do X or he is disposed to we-intend X. Finally, an agent's we-intention to do X amounts to his intention to do his part of X (as his part of X) and some relevant beliefs presupposed in this situation. These intention concepts form a family of concepts which cannot be reduced away (e.g. in terms of personal intentions and mutual beliefs). The irreducibility of the notion of joint intention is also indicated by the fact that a person's we-intention contains implicit reference to a preanalytic notion of joint intention: A person cannot we-intend to do X without we-intending in accordance with and because of the agents' joint intention (plan) to do X together.
Up to now I have used both the agreement-terminology and the plan-terminology to speak about joint intentions. Each of the notions of agreement-making and plan-acceptance would suffice as my central analytic concepts. In fact, as I use these notions, they are equivalent in a clear and obvious sense. Thus, the following thesis, or at least its core idea, can be argued to be acceptable:
(JIP ) Some agents (say A1,...,Ai,...,Am) have formed the (agreement-based) joint intention to perform X if and only if each of them a) has accepted a plan to perform X jointly, b) has communicated this acceptance to the others, and c) because of a) and b) it is a true mutual belief among A1,...,Ai,...,Am that they are jointly committed to performing X and that there is or will be a part or share (requiring at least potential contribution) of X for each agent to perform that he accordingly is or will be committed to performing.
In other words, while the concept of a joint intention is different from the notion of accepting and endorsing a plan to act jointly, still having a joint intention amounts to accepting and endorsing a plan, provided the agents' have - in the simple core cases at least - communicated their acceptance to each other and have become jointly committed to carrying it out (see Tuomela, 1995, Chapter 3 for a discussion).


IV OTHER APPROACHES TO JOINT INTENTION

The study of the notions of joint action and joint intention has not been a popular topic in philosophy. While there is a huge literature on action and intention and related notions in the single-agent case there is relatively little systematic work available concerning the corresponding joint notions. I will below briefly consider all the philosphical approaches that are sufficiently precise to be compared with my above account. These are Bratman's and Searle's accounts. From the approaches published by researchers in DAI I will comment on Cohen's and Levesque's account and briefly also on a paper by Rao, Georgeff, and Sonenberg.
Recently Michael Bratman - who is well-known in AI for his work on single-agent intention - has begun to investigate multi-agent notions. In a recent paper Bratman (1993) analyzes shared intentions in terms of the locutions "We intend to J", where J is a joint action. His final analysis is this (for the two-person case):
We intend to J if and only if
1) a) I intend that we J and b) you intend that we J;
2) a) I intend that we J in accordance with and because of 1)a), 1)b), and meshing subplans of 1)a) and 1)b); you intend that we J in accordance with and because of 1)a), 1)b), and meshing subplans of 1)a) and 1)b).
3) 1) and 2) are common knowledge between us.
I shall not discuss Bratman's notions in detail here but only show that it is not only conceptually different but also classifies cases of shared and joint intention differently from my account. Basically his approach operates with mere personal intentions which concern a joint action. The intentions are interdependent (clause 2)) and mutual knowledge about the agents' intentions is assumed (clause 3)). This conceptually contrasts with my plan-conception which regards the concept of joint intention irreducible to personal intentions and mutual knowledge.
The following argument shows that Bratman's shared intentions need not always be joint intentions in my sense. According to him the following example satisfies his analysis: "You and I arrive at a public basketball court and simply begin, without bothering to assure each other of our intentions, to take turns shooting." According to me, this is best analyzed as a case of mere shared intention based on what I have called a "proper" social norm (one based on mere mutual belief rather than agreement-making), not as a case involving a proper joint intention. Both agents shared a social norm and intend to obey it and there is mutual knowledge about this: this gives the intention to shoot, and Bratman's above conditions for shared intentions are fullfilled, or so Bratman at least assumes. The fulfillment of his clauses does not entail the existence even of an implicit agreement and hence not of a proper joint intention. His analysis accepts the case of shared and suitably interconnected personal intentions - which may be based merely on norm-obeying as above. The basketball example also is a case without strong interpersonal obligations - but there is no joint intention in my plan-sense either. So the case is compatible with my analysis and also serves to show that there are cases of shared intention in Bratman's sense which fail to be cases of joint intention in my sense. I claim that my analysis is more informative as an analysis of joint intention in that it discriminates better between such diverse examples as personal norm-obeying and "true joining of wills" which takes place in in the case of agreement-based joint intention. However, Bratman's notion of shared intention may be of interest in its own right.1
Let me next consider the account by the philosopher John Searle (1990).2 He adopts a view of collective intentions which is rather close to my view (which I have presented and elaborated - with some changes and corrections - in a number of publications since the late 1970's). His view of the content of we-intention resembles what I have said about the satisfaction conditions of we-intention. According to him the content (satisfaction conditions) of a we-intention is means-relative: Ai intends to perform joint action X by means of his single-agent action Xi. But this idea holds true in my account as well, for the parts Xj, j = 1,...,m, are means to the m-agent joint action X. Searle gives the following formula for the satisfaction conditions of a we-intention (intention-in-action, i.a.), B being the joint action of mixing a sauce (p. 414):
i.a. collective B by means of singular A (this i.a. causes: A stirred, causes: B mixed).
This is not quite right, because it ignores that a participant's part-performance is dependent on the other participants' part performances. But corrected for this, Searle's account - when corrected also for lack of the resources of "conceptual action generation"- more or less amounts to my view. We then get basically this:
Ai has a we-intention to perform a joint action X by means of his performing his part Xi (as his part of X), presuming that the other participants perform their parts of X; this we-intention has as its conditions of satisfaction that this very we-intention (causally and/or conceptually) generates Ai's performing his part Xi, which, presuming the other participants' performances of their parts, (causally and/or conceptually) generates an intentional performance of X by the participants.
Let me now briefly consider how researchers in DAI have analyzed joint intention. Actually the only published work I have seen comes from two sources. First, Cohen and Levesque and their co-workers have presented a detailed full-blown theory of intention and action both for the single-agent and the multi-agent case. Their theory is surely a very fine achievement, although I do have some criticisms concerning its details. Nobody working on joint intention within DAI or in any other field can afford not to study the papers of these authors. I shall below sketch in informal terms - used by Cohen and Levesque themselves - some basic features of their complicated theory and I will also present some critical remarks against it.
In Cohen and Levesque (1991) the following definition of joint intention is given:
Definition. A team of agents jointly intends , relative to some escape condition, to do an action, iff the members have a joint persistent goal relative to that condition of their having done that action and, moreover, having done it mutually believing throughout that they were doing it.
The notion of a joint persistent goal used here is analyzed as follows:
Definition : A team of agents have a joint persistent goal relative to q to achieve p just in case
1) they mutually believe that p is currently false;
2) they mutually know they all want p to eventually be true;
3) it is true (and mutual knowledge) that until they come to mutually believe either that p is true, that p will never be true, or that q is false, they will continue to mutually believe that they each have p as a weak achievement goal relative to q and with respect to the team.
The notion of a weak achievement goal is this:
Definition : An agent has a weak achievement goal relative to q and with respect to a team to bring about p if either or these conditions holds:
i) The agent has a normal achievement goal to bring about p, that is, the agent does not yet believe that p is true and has p eventually being true as a goal.
ii) The agent believes that p is true, will never be true, or is irrelevant (that is q is false), but has as a goal that the status of p be mutually believed by all the team members.

The conditions about the existence of joint action and joint intention opportunities mentioned above seem to be correct. Thus, surely when having a joint goal or intention that p, the participants are assumed not to mutually believe that p or that necessarily -p or that -q. (These conditions are also covered in my account, not as conditions in the analysis itself but as presuppositions of analysis.) What is especially valuable in these authors' analysis is that they show what role divergent "private" beliefs about the situation had by the participants may play here. For instance, such a private belief when having the content that the joint action after all faces insurmountable obstacles may lead a participant to switch his original goal of performing his part of the joint action into the goal - or, as we should rather say, obligation - of creating the mutual belief among the participants that the goal of the joint action is not attainable.
However, as seen from the above summary, this account does not refer to either explicit or implicit agreement (not even to believed agreement); nor is obligation or any other normative notion involved in the analysans; finally, communication is not "officially" required in the Cohen-Levesque theory. In my account presented in Section V these features, in contrast, play a central role, and suggest to me that my analysandum may be stronger than in these authors' account. Furthermore, in their theory, joint intention and joint commitment are basically analyzed in terms of goals (chosen, consistent desires) and mutual (and other) beliefs. This analysis problematically reduces away intentions - unless their notion of choice here is a necessarily intentional notion. Their theory also involves the possibility that noncommunicating robots or any similar mechanical devices having beliefs or belief-like states but lacking normative conceptual resources could have joint intentions and commitments. From the point of view of our theory this is implausible at least if full-blown joint intention is meant. Mutual belief surely is necessary in this context, but the normative notion of agreement (at least belief in agreement) is also needed. It seems that cases of mere norm-following with rigid normative constraints for the beginning and end of the instances of norm-following would qualify as joint intentions in the present authors' sense (cf. the Finnish ritual of lighting candles in the evening of the Independence Day).3 From a normative point of view, it can be said that the analysans in the account of Levesque et al., because involving no normative notions, seems to me not to be fully able to account for the notion of joint commitment, no matter how well it may capture the descriptive aspects of commitment.
I have argued above (and in other works) that conceptually the notions joint intention, joint commitment, and joint action can only be based on the normative (viz. joint-commitment entailing) notion of agreement (or believed agreement), implicit or explicit, and that, on factual grounds, recurrently successful joint action would be a surprising phenomenon in cases of several agents trying to act jointly merely on mutual belief, without agreement, about what they and the others should and will do in a joint project. This suggests to me that the Cohen-Levesque approach, no matter how detailed and precise, is not fully adequate.
In a recent paper the DAI-researchers Rao, Georgeff, and Sonenberg (1992) present a technical account of joint intention. While in the Cohen-Levesque logical semantics possible worlds are time-lines the present authors take possible worlds to be trees. This gives more logical flexibility but also complicates the analysis. However, I cannot here go into technical questions. As to joint intention in the present approach it is a primitive notion which is connected to plans and personal attitudes (intentions, goals, and beliefs).4 While these authors' work surely is interesting from a logical point of view, it - perhaps because of its preliminary nature - adds relatively little to the conceptual features of joint intention.


V CONCLUSION

I hope the above presentation has shown that work on jointness notions is in full swing both in philosophy and in DAI and also that there is genuine interaction between the research done in these two different fields. One would expect that in the near future - possibly in a few years - most of the central social notions mentioned in Section I of this paper will a topic of interest and - hopefully - of actual work within DAI. There is one good reason for this expectation: As work on multi-agent notions gets further it simply has to - due to conceptual-theoretical reasons - deal with a number of central, intertwined notions that so far have not been studied by DAI-researchers. The BDI-paradigm will have to be expanded to deal with - for instance - agreement, joint decision, social power, communication, and various institutional social concepts.




Notes
1) In Tuomela (1994) I argue that Bratman's analysis - which is not based on the plan-sense of joint intention - is too weak even as an analysis of shared intention. The basic point of the argument is that the agent's intentions in clause 1) of his analysis must be responsive to the mutual knowledge of clause 3). But when corrected for this, Bratman's analysis becomes formally equivalent to the analysis of the notion of an intended collective goal that I advocate in the mentioned book. This notion is wider than the plan-based notion of joint intention and is also reducible to individual intentions and beliefs. It has important applications to theory of social action.

2) Searle (1990) presents a criticism of the account of we-intention presented in Tuomela and Miller (1988). I argue in Tuomela (1995), in a long footnote to Chapter 3, that Searle has badly misunderstood our position and that therefore his criticism has very little to do with what our view is. He believes that our position is that joint intentions can be reductively defined in terms of personal intentions and mutual beliefs. As seen above, this is not my view. I take Searle's criticism against that reductive position to be right.

3) Castelfranchi et al. (1993) give another kind of counterexample. Two researchers (or, if you prefer, research teams) have the intention to find a vaccine against AIDS. They argue that this shared intention amounts to a joint intention in the sense of Cohen and Levesque. However, as the researchers compete with each other their shared intention cannot be a joint intention.

4) Rao et al. give their definition on joint intention on p. 68. Let me try to give its basic content verbally: The members of a collection of agents jointly intend that something p if and only if everyone in the collection intends that p and there is a mutual belief among them that they jointly intend that p. Thus we can see that the notion of joint intention occurs irreducibly in the content of the agents' mutual belief. This results seems to me to contradict these authors' statement on p. 60 according to which "all joint propositional attitudes are defined in terms of the individual propositional attitudes".
Rao et al. regard my (and my coauthor Miller's) work as reductive: "Tuomela and Miller [1988] and Grosz and Sidner [1990] regard a joint intention as definable in terms of the conjunction of individual intentions together with mutual beliefs about the intentions of other individuals". As seen, this claim does not hold true of our approach while it may apply to the view of Grosz and Sidner.



References

Belnap, N. (1991), 'Backwards and Forwards in the Modal Logic of Agency', Philosophy and Phenomenological Research LI , 777-807
Bratman, M. (1987), Intention, Plans, and Practical Reasoning , Harvard University Press, Cambridge
Bratman, M. (1992), 'Shared Cooperative Activity,' The Philosophical Review 101 , 327-341
Bratman, M. (1993), 'Shared Intention', Ethics 104 , 97-113
Bratman, M. (1994), 'Shared Intention and Mutual Obligation', Cahiers d'epistemologie, no 9319, Universite du Quebec a Montreal
Castelfranchi, C., Cesta, A., Conte, R., Miceli, M. (1993), 'Foundations for Interaction: The Dependence Theory', in Torasso, P. (ed.), Advances in Artificial Intelligence. Third Congress of the Italian Association for Artificial Intelligence , Springer-Verlag, Heidelberg, pp. 59-64
Cohen, P. and Levesque, H., 1991, 'Teamwork,' Noûs 35 , 487-512
Cohen, P., Morgan, J., and Pollack, M. (eds.) (1990), Intentions in Communication , A Bradford Book, The MIT Press, Cambridge, Mass.
Gilbert, M., 1989, On Social Facts , Routledge, London
Grice, P. (1957), 'Meaning', The Philosophical Review 66 , 377-388
Grosz, B. and Sidner, C. (1990), 'Plans for Discourse'. In Cohen, P. et al. (1990), 417-444
Levesque, H., Cohen, P., and Nunes, J. (1990), 'On Acting Together,' in Proceedings of Eighth National Conference on Artificial Intelligence, vol I , The MIT Press, Massachusetts, 94-99
Pettit, P. (1993), The Common Mind , Oxford University Press, New York
Rao, A., Georgeff, M., and Sonenberg, E. (1992), 'Social Plans: A Preliminary Report', in Werner, E. and Demazeau, Y. (eds.), Decentralized A.I. 3 , North-Holland, Amsterdam, 57-76
Sandu, G. and Tuomela, R., 1994, 'Joint Action and Group Action Made Precise,' Synthese (forthcoming)
Searle, J. (1990): 'Collective Intentions and Actions,' in Cohen, P. et al. (1990), 401-415
Tuomela, R., 1984, A Theory of Social Action , Synthese Library, Reidel Publishing Company, Dordrecht and Boston
Tuomela, R., 1991, 'We Will Do It: An Analysis of Group-Intentions,' Philosophy and Phenomenological Research LI , 249-277
Tuomela, R., 1993a, 'What is Cooperation?,' Erkenntnis 38 , 87-101
Tuomela, R., 1993b, 'What are joint intentions?', in Casati, R. and White, G. (eds.), Philosophy and the Cognitive Sciences , Austrian Ludwig Wittgenstein Society, Kirchberg, pp. 543-547
Tuomela, R. (1994), Cooperation: A Philosophical Study , book manuscript
Tuomela, R. (1995), The Importance of Us: A Philosophical Study of Basic Social Notions , Stanford University Press (forthcoming)
Tuomela, R. and Miller, K. 1988, 'We-Intentions,' Philosophical Studies 53 , 115-137
Tuomela, R. and Sandu, G., 1994, 'On Seeing to It that Something Is the Case,' in P. Humphreys (ed.), Patrick Suppes: Scientific Philosopher, Vol. 3 , Kluwer, pp. 193-221