intention maintenance, intention, desire, semantic constraints, semantic constraint, belief revision, formalization, M. P. Georgeff, rational agents, A. S. Rao, M. Pollack, Knowledge Representation and Reasoning, Technical Report, Morgan Kaufmann Publishers, International Conference, K. Konolige, agent-oriented, axiomatization, Levesque, intentions, truth function, non-determinism, maintenance, Artificial Intelligence Institute, Artificial Intelligence
The Semantics of Intention Maintenance for Rational Agents
Michael P. Georgeffand Anand S. Rao Australian artificial intelligence
Institute Level 6, 171 La Trobe Street, Melbourne Victoria 3000, Australia
Abstract The specification, design, and verification of agent-oriented systems depends upon having clear, intuitive formalisms for reasoning about the properties of such systems. In this paper, we consider agents whose state comprises the three mental attitudes of belief, desire, and intention. While the static relationships among these entities has had considerable attention, the manner in which these entities change over time has not been formalized rigourously. By considering some simple examples, we show that the formalization of some of these intuitions is problematic. We then examine these intuitions from a possible-worlds perspective and formally describe the dynamics of intention maintenance in the context of changing beliefs and desires. To solve the problems identified in the examples, and to properly capture our semantic intuitions about intention maintenance, we extend the standard logics by introducing forms for only modalities of belief, desire, and intention, along the lines of Levesque's only believe operator. This allows us to formalize the process of intention maintenance. We conclude by comparing our work with other related work. 1 Introduction Agent-oriented systems are finding increasing application in the commercial world. One of the most successful of agent architectures is that based around the notions of belief, desire, and intention (BDI) [2; 3; 4; 8; 14], representing respectively the informative, motivational, and decision components of the agent. Such systems have been applied to a wide range of large-scale applications, including Air Traffic Control
, telecommunications network management, business process management, and simulation. Within such systems, intentions play an essential role. First, prior intentions pose problems for further deliberation; in AI terms, they specify the goals (ends) for further means-ends analysis. Second, prior intentions constrain the deliberation process because they rule out options that conflict with existing intentions. Under this view, the deliberation process is a continuous
resource-bounded activity rather than a one-off exhaustive decision-theoretic analysis . The critical element in this view of practical reasoning is that the adoption of an intention entails some form of commitment to that intention. That is, intentions only have value if they are maintained from one time point to the next--if they are not so maintained, they can establish neither the goals for further deliberation nor the basis for ruling out conflicting options. However, the specification, design, and verification of such systems depends on being able to semantically model these agents and formally describe the process of intention maintenance and the resulting agent behaviour. A number of formalisms that provide the semantics of intention and its relation to the other attitudes, such as belief and desire, have been proposed in 'the literature. In providing these formalisms, various possible static relationships among belief, desire and intention have been considered. In essence, most of these reflect the intuition that one only adopts an intention to an action or proposition that is (i) desired and (ii) believed to be possible. The variations on this basic intuition concern certain special cases that one may or may not consider important, dependent on the purpose of the formalization. In addition, some authors have proposed certain axioms to capture the dynamic relationships among these attitudes, particularly those concerning the maintenance of intentions. The intuition here is that an intention should be maintained as long as the object of the intention (i) is continued to be desired, (ii) is continued to be believed possible, and (iii) has not yet been achieved.1 Unfortunately, the translation of this condition into formal axioms of intention maintenance is more problematic than it first appears. In fact, it turns out that to express these dynamic properties of intention maintenance requires a more expressive logic than has been considered in the literature so far. In this paper, we base our approach on a possibleworlds model developed previously [10; 12; 13]. However, the results we obtain apply more generally. *In this paper we make the simplification that the object of the intention, once achieved, is no longer desired.
704 DISTRIBUTED AI
2 The Problem
3 Informal Model
For simplicity, let us consider the relationship between intentions and beliefs only. From the discussion above, one would expect the formalization of the maintenance condition for intentions to take something like the following form:
The formal and informal models of our BDI agents have been discussed elsewhere [10; 12; 13]. In this section, we briefly describe our model and then motivate the static and dynamic relationships between different entities within the model.
where is true at the next time point. Consider a situation in which John intends to go to the beach. From the above axiom, John will maintain this intention as long as he believes it to be achievable.2 If, or when, John discovers that it is not possible to go to the beach, this intention can be dropped (and, indeed, the static constraints would force it to be dropped). This is just what we want. However, let's assume that John also believes it is possible to fly to London, but has no intention of doing so. Because we have \NTEND(X(go-to-bcach)) we also have (under a possible worlds model) the disjunctive intention INTEND(X(go-to-beach V go-to-London)). Now, when it turns out that visiting the beach is impossible, the intention towards visiting the beach will be duly dropped. But, unfortunately, the intention towards the disjunction (go-to-beach V go-to-London) will be maintained (as the disjunct remains a possibility). From application of the static constraints, it can then be deduced that John, at the next time point, will intend to fly to London. In other words, John will be forced to adopt as new intentions any beliefs about the future he still holds! A similar problem arises in the the following situation. John intends to obtain milk from the milk bar and cereal from the supermarket. He goes to the milk bar, sees that it is closed, and thus abandons the intention of obtaining milk. As a result John also gives up his intention to have milk and cereal. However, if intentions are closed under conjunction--as they are in a possible worlds model-- intending to have milk and cereal implies an intention to have milk and an intention to have cereal. While the former two can no longer hold, using the above axiom of intention maintenance, the intention to have cereal would be (incorrectly) maintained. Noting similar problems, Konolige and Pollack also considered closure under conjunction to be a problem for intentions, although in relation to static rather than dynamic properties. Their solution involves representationalist approach to the modelling of intentions . But what is the real problem here? Is it simply that we do not want closure under conjunction, or is our simple axiomatization just not properly capturing our intuitions? While some of the undesirable symptoms of the problem are clear, the cause is not. The approach we adopt here is to go back to our semantic model
and understand what was really intended by the conditions of intention maintenance, and to develop axioms that properly reflect our semantic intuitions.
Our semantic model consists of sets of possible worlds where each possible world is a branching tree structure with a single past. A particular index within a possible world is called a time point or situation. The branches within a tree structure represent different courses of action or execution path
s. We model the beliefs of the agent by a set of such possible worlds, called the beliefaccessible worlds of the agent. Similarly, we model the desires and intentions of the agent by a set of desireaccessible worlds and a set of intention-accessible worlds, respectively. The different belief-accessible worlds represent the agent's lack of knowledge
or chance inherent in the environment; that is, as far as the agent knows, the actual world could be any one of the belief-accessible worlds. Within each belief world, each path represents the options or choice of action available to the agent. Corresponding to each belief-accessible world is a desire-accessible world and an intention-accessible world.3 These represent, respectively, the desires and intentions of the agent with respect to that particular belief world (that is, the desires and intentions the agent would have if that world was known to be the actual world). Each path within the desire-accessible world represents an execution path that the agent wants to achieve (or is content to achieve), and each path within an intention-accessible world represents an execution path that the agent has decided upon (one of which, in the context of our earlier discussion, the agent is committed to bringing about). Now consider the static structural relationships among such a triple of belief-desire-intention worlds. While, for any such triple, we place no constraints on the relationship between the paths believed possible and the desired paths, we require that the intention paths be a subset of both (see Figure 1). This reflects the intuition that one will only intend a course of action that is both believed possible and desired.4 But what happens now as we move from one time point to the next (from t to v in world w as shown in Figure 1)? The basic intuition is that, provided the agent's beliefs and desires are not significantly changed, the agent's intentions will be maintained. More specifically, for any triple of belief-desire-intention worlds, we would like to retain any existing intention path provided it was still both believed possible and desired. Any intention path that was no longer believed possible, or was no longer desired, would be pruned off the intention structure. Any new belief paths, i.e., new opportunities (shown as a dotted path with r true in the future in
2Clearly, we need to add additional conditions to account for changing desires and the successful achievement of John's intentions. However, for simplicity, we do not consider these situations here.
3 We elsewhere  consider the more general case where we relax the requirement for such a one-to-one correspondence. 4For discussion of this point, see our earlier work [10; 12].
GEORGEFF AND RAO 705
Figure 1: An example of Belief, Desire, and Intention Revision 706 DISTRIBUTED Al
the semantic constraints SCI and SC2. Axiom Al states
that any intended execution path must be believed to be
possible (that is, must be believed to be an option for
the agent). Axiom A2 states that any inevitable belief will be intended.6 Axioms A3 and A4, resulting from
the constraints SCI and SC3, state that any path that
is intended must be desired and any inevitable desire will
be intended. Axiom A5, resulting from constraints SCI
and SC4, states that at least one of the desired execution
paths is believed achievable.
To preserve the mapping to decision tree
s [ll], we
make the following deterministic world assumption. This
assumption requires for a given model, and all world time
point pairs, that
then 6 =
b', where L is the truth assignment function. Intuitively,
this means that there is no additional non-determinism
beyond that represented by different belief worlds. In
other words, the real world is deterministic; any per-
ceived non-determinism results from an agent's lack of
knowledge of the world. Similar assumptions hold for desire- and intention-accessible worlds.7
We refer to the above axiomatization together with
the axioms relating intention and action (see our earlier
work[l2] for details) as the BDI-modal system. Other
variations to this axiomatization can be obtained by al-
lowing the total 1-1 mappings /, g and h to be partial,
which account for the cases that have been referred to
as realism , weak-realism , and strong-realism .
Different structural relationships can also be adopted
among B-, V-, and Z-accessible worlds to obtain further
variations in the axiomatizations.
It turns out, however, that under all these variants
we need some additional expressive power to capture
the notion of intention maintenance discussed above. To
achieve this, we now extend the language BDICTL* by
introducing only forms of the modalities for beliefs, de-
sires, and intentions. Intuitively, if an agent only intends
a formula then is true in all the intention-accessible
worlds and the set of intention-accessible worlds includes
all worlds where is true.
The definition of INTEND ) includes only the if part of the definition above. It is important to note that, whereas the operator INTEND is closed under conjunction, OINTEND is not. That is, we have the following theorem: Theorem 1 The following statements are true of the OINTEND operator.
6As discussed in our previous work  these axioms can be weakened by adopting alternative semantics constraints to that of SCI and SC2. 7The mappings /, g, and h are uniquely determined by the truth function assignment L, given the assumption of a deterministic world.
The proof is straightforward . For example, the above properties of the only intend modality entail that, if John only intends having milk and cereal for breakfast, he will not necessarily also only intend having milk and only intend having cereal. Similarly, if John only intends to go to the beach, he will not necessarily also only intend to go to the beach or only to go to London.
5 Maintenance of Intentions
Now let us consider the problem of an agent maintaining its intentions as the environment changes. Our aim is to specify semantic constraints on our models that will determine how the model changes from one time point to another. In so doing, we will treat the processes of belief and desire revision as given and consider how these processes determine intention revision. Let us assume that the agent revises its beliefs using some well-known belief revision or update procedure [l]. For the purpose of this paper, we assume that the nondeterminism (chance) inherent in the beliefs of the agent remains constant over time. Intuitively, this corresponds to an agent believing it is in one of a number of possible worlds, its beliefs about which can change over time, but about which it can never get sufficient information to eliminate any from consideration. It may, for example, discover that, for any particular possible world, it has different options than previously believed, but will not be able to reduce the uncertainty concerning which possible world it is actually in. Under this assumption, at the semantic level the belief revision function is a total 1-1 mapping; that is, the belief revision process maps each old belief world into a corresponding new belief world. The propositions that hold in that new belief world may be quite different from those that held in the previous belief world, but no new belief worlds are introduced nor old ones deleted. Although this seems restrictive, the assumption can be relaxed without too much difficulty by removing the semantic constraint SCI on the functions /, g, and h. However, for the purposes of this paper, this unnecessarily complicates the picture. We therefore postulate a belief revision function which maps each belief-accessible world to its revised belief-accessible world. More formally, we have:
Definition 1 For each world w and time t the belief revision function is a mapping from the set of beliefaccessible worlds at t to the set of belief-accessible worlds at the next instant v. Formally,
We postulate similar desire revision and intention re-
vision functions for each world w and time point t, de-
and , respectively. Figure 2 shows
the various functions involved in the revision process.
Each solid circle represents a world which is a branch-
ing tree structure. The set of belief-accessible worlds
at world w and time t has a total 1-1 mapping to
its corresponding desire-accessible (denoted by , I and
intention-accessible worlds (denoted by ). The belief
revision function maps each world in to its corre-
sponding world in and similarly for the desire and
intention revision functions. The functions ,
GEORGEFF AND RAO 707
708 DISTRIBUTED Al
5.2 Desire Revision The same belief filtration principle applies for desires as well. In other words, when an agent revises its intentions it should ensure that the new intentions are compatible with its new desires. Semantically, we therefore impose the constraint that the intention-accessible worlds are filtered through the corresponding revised desire-accessible worlds to obtain new intention-accessible worlds. Of course, we want, our intentions to be compatible with both beliefs and desires. This results in the new intention paths being the intersection of new believed paths, new desired paths, and old intention paths.
Finally, it is worth considering two variations to the above model of intention maintenance: (a) what happens if the new belief-accessible world contains a new option (e.g., a path ending in the proposition r) that was not present in the previous time point; and (b) what happens if the filtered intention-accessible world has no future options (e.g., if the original intention-accessible world did not have the path ending only in q). In the first case, intention maintenance will ensure the stability of intentions but does not allow the exploitation of new opportunities. As a result, any additional options that are part of the revised belief-accessible world will not be included in the corresponding new intentionaccessible world. This is exactly what one wants for intention maintenance. However, this does not mean that new options can never be considered--an agent with sufficient computational resources may reconsider its intentions in the light of new opportunities. This can be modelled as a separate process following the above fdtering process. In the second case, no intentions will be maintained and the agent has no choice but to reconsider his available options. That is, the agent would have to deliberate anew [ll] to derive new intention-accessible worlds from its current belief- and desire-accessible worlds. Similar results can be expected to hold when we relax the constraints that the revision functions be total 1-1 mappings (together with the semantic constraint SCI). However, this goes beyond the scope of this paper.
6 Comparison and Conclusion Cohen and Levesque  define the notion of intention in terms of the other entities, such as beliefs, goals, persistent goals, and actions. In their formalism, an agent has a persistent goal or PGOAL if and only if the agent currently believes has the goal to eventually make true, and maintains this goal until it either comes to believe in , or comes to believe that is impossible. PGOAL is closed under conjunction except in the special case where the agent already believes that one of the conjuncts is true or when the conjuncts hold at different time points. As neither example given in Section 2 is one of these special cases, the problems identified there are also exhibited in Cohen and Levesque's theory. Similarly, PGOAL is closed under disjunction except in very special circumstances
. One could rectify the problems by adopting a similar approach to that used here. As mentioned earlier, Konolige and Pollack  claim that Normal Modal Logics (NML) are not suitable for modelling intentions. They introduce a model of intentions that has two components: "possible worlds that represent possible future courses of events, and cognitive structures
, a representation of the mental state components of an agent" . They define a scenario for a proposition as the set of worlds in W that make true, denoted by . A n agent intends iff the set of scenarios for is identical to the set of scenarios for any intention in the cognitive structure of the agent. This has an interesting correlation with our definition of OINTEND, if one considers each of their intention worlds as a path in our branching tree structures. The primary difference between the
GE0R6EFF AND RA0 709
two approaches being that Konolige and Pollack follow a syntactic or representationalist approach and we follow a semantic approach. As a consequence, in their approach one has to explicitly conjoin formulas in the set of intentions given by the cognitive structure. Our semantic approach makes this unnecessary. Moreover, and perhaps more importantly, the semantic approach allows us to address the cause of the problem, not its symptoms. Konolige and Pollack do not address the issue of belief and intention revision but do extend the notion of cognitive structures in terms of the plans of an agent. In this paper, we have explored the role of the only modalities in intention revision, but have remained silent on the important notion of plans . The only modality was introduced by Levesque  in the context of beliefs and non-monotonic reasoning to capture the notions of stable sets in autoepistemic logic on a semantic basis. We have used the same concept for all the mental attitudes of the agent to give semantic characterizations of intention revision. The primary contribution of this paper has been to lay out a semantic story of intention maintenance in the context of changing beliefs and desires. By introducing the only modalities to exactly specify paths of execution, we have also been able to provide a sound axiomatization of the intention maintenance process. Of course, considerable work remains to be done. The completeness of the axiomatization needs further investigation. In addition, the restrictive conditions on the correspondence functions relating beliefs, desires, and intentions need to be removed and the proofs redone in this context. Finally, we need to show clearly how all this fits equally well within a decision-theoretic framework. Acknowledgements: This research was supported by the Cooperative Research Centre for Intelligent Decision Systems under the Australian Government
's Cooperative Research Centres Program. References  C. Alchourron, P. Gardenfors, and D. Makinson. On the logic of theory change: Partial meet contraction functions and their associated revision functions. Journal of Symbolic Logic, 50:510-530, 1985.  M. E. Bratman. Intentions, Plans, and Practical Reason. Harvard University
Press, Cambridge, MA
, 1987.  M. E. Bratman, D. Israel, and M. E. Pollack. Plans and resource-bounded practical reasoning. Computational Intelligence, 4:349-355, 1988.  P. R. Cohen and H. J. Levesque. Intention is choice with commitment. Artificial Intelligence, 42(3), 1990.  J. Doyle. A truth maintenance system
. Artificial Intelligence, 12:231-272,1979.  E. A. Emerson. Temporal and modal logic. In J. van Leeuwen
, editor, Handbook of Theoretical computer science
: Volume B, Formal Models and Semantics, pages 995-1072. Elsevier Science Publishers and MIT Press, Amsterdam and Cambridge, MA, 1990.
 M. P. Georgeff and A. S. Rao. The semantics of intention maintenance for rational agents. Technical Report
57, Australian Artificial Intelligence Institute, Melbourne, Australia
, 1995.  K. Konolige and M. Pollack. A representationalist theory of intention. In Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence (IJCAI-93), Chamberey, France, 1993.  H. J. Levesque. All I know: A study in autoepistemic logic. Artificial Intelligence, 42(3), 1990.  A. S. Rao and M. P. Georgeff. Asymmetry thesis and side-effect problems in linear time and branching time intention logics. In Proceedings of the Twelfth International Joint Conference on Artificial Intelligence (IJCAI-9I), Sydney, Australia
, 1991.  A. S. Rao and M. P. Georgeff. Deliberation and its role in the formation of intentions. In Proceedings of the Seventh Conference on Uncertainty in Artificial Intelligence (UAI-9I). Morgan Kaufmann
Publishers, San Mateo, CA, 1991.  A. S. Rao and M. P. Georgeff. Modelling rational agents within a BDI-architecture. In J. Allen, R. Fikes, and E. Sandewall, editors, Proceedings of the Second International Conference
on Principles of knowledge representation
and Reasoning. Morgan Kaufmann Publishers, San Mateo, CA, 1991.  A. S. Rao and M. P. Georgeff. A model-theoretic approach to the verification of situated reasoning systems. In Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence (IJCAJ98), Chamberey, France, 1993.  A. S. Rao and M. P. Georgeff. BD1 Agents: From theory to practice. In Proceedings of the International Conference on Multi-Agent Systems (1CMAS-95), San Francisco
, USA, June, 1995.
710 DISTRIBUTED Al
AS Rao, MP Georgeff