Social pathologies of adaptive agents

摘要

Multiagent systems do not always behave as their designers intend. In some cases, a simple flaw in design or implementation causes sub-optimal behavior, and such flaws are often easy to remedy. In other cases, sub-optimal behavior is caused by fundamental characteristics of system structure or problem representation, and these flaws can be difficult to understand and avoid. Such cases are often called pathologies. Several pathologies of single-agent systems have been studied extensively. For example, game-tree search algorithms exhibit pathological behaviors in certain, limited classes of games (Nau 1983, Kaindl 1988). Similarly, many induction algorithms exhibit overfitting, a tendency to produce models that contain excessive structure (Jensen and Cohen 2001, Oates and Jensen 1997). Our research focuses on a particular class of pathologies of multiagent systems - social pathologies of adaptive agents. Such pathologies are adaptations of individual agents that are locally beneficial but that degrade the performance of the overall system. Social pathologies are particularly troublesome because they contradict a central assumption of multiagent systems - that local performance improvements by one or more agents will lead to system-wide performance improvements. Below, we define social pathologies and analyze the conditions under which they arise. We provide several examples of social pathologies and draw examples from diverse multiagent systems. Finally, we briefly discuss how understanding social pathologies can improve our knowledge of other aspects of multiagent system behavior.