Adaptive model of multi-objective agent behavior in real-time systems

Abstract

Agents trying to reach their goals in dynamical environments need to be adaptive. Adaptation rules can conflict and their combinations can complicate multi-agent systems development. Agent programs are becoming mostly problem-oriented and high-coupled, and that prevents the reuse of developed programs and their components. High-level generic planning algorithms can be used to adapt agent behavior to environment changes, but they also have high computational complexity, which limits their usefulness in real-time application. This paper considers a generic agent model that combines planning algorithms with utility theory to reach rational adaptive behavior, with acceptable performance and low component coupling. In the proposed model each agent has a dynamical set of tactical objectives associated with objects in environment. For each kind of objective the agent uses an objective-specific planning algorithm. To handle events from the environment the agent generates new objectives with a decision tree. Each objective has a dynamic priority calculated by heuristic function. The suggested model may improve performance in real-time environments by decreasing computational complexity. The effectiveness of the model is shown on mini-game example.