6. Conclusion

16 disjoint. It overcomes the hypothesis of existence of a precise final goal in goal-regression systems. Motivations have been shown to be useful in guiding the behavior of a mobile robot in variously configured indoor environments. Experimental results provide evidence that correlation does exist between motivational profile and environment, in order to get optimal performances of the agent. Our ongoing research involves (1) mathematically characterizing the optimality of some motivational profiles in some environments, by finding an invariant in the pseudo decision tree that is continuously generated, and (2) embedding into the agent the knowledge necessary to recognize the type of its environment and change its motivational profile accordingly. Future research includes reversing the method, i.e., a posteriori rebuilding the motivational profile of another agent, based on its perceived behavior. Ackowledgements We would like to thank all our colleagues of the AIbots project for many fruitful discussions. Dan Shapiro, Marcel Schoppers, Aaron Sloman and anonymous reviewers made various contributions. 15 states are looked for in future possible worlds, a partial order on test-action pairs is generated to avoid reaching undesirable futures, then the agent indefinitely reacts to its environment in the way prescribed by polling this structure [Schoppers 87] [Nilsson 89]. The first approach provides useful tools for deliberating before acting but such agents lack reactivity which is necessary in dynamic and even unpredictable environments-generating a plan is an NP-complete activity [Chapman 87], which essentially means that a search has to be performed. The second approach provides a reactive architecture for a given environment but lacks reconfiguration capability and avoids representing deliberation, which limits the agent to perform already planned tasks and thus forces the agent designer to rebuild an entirely new agent for each new environment. When compared to ours, both of these approaches assume that the agent integrates in advance a precise and unique description of its final goal. But real agents in unpredictable environments have to cope with a stream of incoming goals, only the initial situation of the planning problem is precisely known when action starts. • A second, non-traditional, trend ("new A.I.") proposes a situated view of the agent in its environment. (1) Survival. An agent can be considered as a society of small reactive agents, the aggregation of which produces a ("emergent") behavior that ensures the actual survival of the agent and the achievement of given goals, so that agents always Do The Right …