Goal-based personalities in lifelike agents realized with hybrid planning techniques

Paola Rizzo, Maria Miceli, Amedeo Cesta

[paola | maria | amedeo @pscs2.irmkant.rm.cnr.it]

IP-CNR
(Institute of Psychology of the National Research Council)
Viale Marx 15
I-00137 Roma, Italy







Personality, intended as a coherent pattern of behavior that characterizes each agent and makes it clearly distinguishable from others, is generally considered a crucial feature to make an agent more lifelike and appealing to the user.

Our model of personality is focused on the motivational aspects of agents. It is based on the concept of "General Goals" (G-GOALS), that represent abstract states or events an agent recurrently attempts to achieve and maintain through its behavior, and on goal-based preferences over actions and plans, determined by how the latter can affect the G-GOALS through their "side-effects". The model has been instantiated into the behavior of help-giving, thereby producing personalities like the "altruist", the "selfish", the "normative", the "spiteful", and the "suspicious", which are characterized by distinct clusters of G-GOALS having different priorities.

The model is realized with a software architecture aimed at taking advantage of both the useful properties of classical planners, and the ability of reactive planners to allow the designer a deep control over the agents' behaviors. Therefore, it integrates the PRODIGY[VCP+95] generative planner and the RAP[Fir89] reactive planner. PRODIGY has been enabled to automatically produce personality-biased plans according to the goal-based model, while RAP allows the designer to customize operators and plans and to hand-code further plans as needed.

An agent can interact with a user by letting PRODIGY pursue its typical G-GOALS and execute its preferred actions, possibly asking PRODIGY to generate new behaviors at run time. In order to make this possible, several research issues have been addressed, concerning how to (1) represent G-GOALS and actions with side-effects in PRODIGY; (2) define suitable heuristics for making PRODIGY automatically compute goal-based preferences over actions, in order to produce personality-biased plans; (3) represent G-GOALS having different priorities and hand-code personality-biased plans in RAP; (4) map PRODIGY operators and plans into plans that can be performed by RAP; (5) close the loop between PRODIGY and RAP at execution time; (6) take social aspects into account when an agent performs behaviors while interacting with the user.

The goal-based model of personality, and the integrated planning architecture, have been concretely used for realizing agents that interact with the user by pursuing goals and displaying behaviors that are typical of different "helping personalities". The model, the integrated planning architecture, and the results obtained so far are described in more detail in [RVMC99], [R98], and [RVMC97], some of which are available on the web.
 
 

References

[Fir89] R. J. Firby. Adaptive Execution in Complex Dynamic Domains. PhD thesis, Yale University, 1989. Technical Report YALEU/CSD/RR 672.

[R98] Rizzo P. Personalities in believable agents: A goal-based model and its realization with an integrated planning architecture. PhD thesis, University of Torino, Italy, 1998.

[RVMC97] Rizzo P., Veloso M. Miceli M., M., and Cesta A. Personality-Driven Social Behaviors in Believable Agents. AAAI 1997 Fall Symposium on "Socially Intelligent Agents", AAAI Press Technical Report FS-97-02, pp. 109-114.

[RVMC99] Rizzo P., Veloso M. M., Miceli M., and Cesta A. Goal-based personalities and social behaviors in believable agents. Applied Artificial Intelligence, special issue on "Socially Intelligent Agents", in press.

[VCP+95] M. M. Veloso, J. Carbonell, A. Perez, D. Borrajo, E. Fink, and J. Blythe. Integrating Planning and Learning: The prodigy Architecture. Journal of Experimental and Theoretical Artificial Intelligence, 7:81--120, 1995.