Constructing optimal policies for agents with constrained architectures

Optimal behavior is a very desirable property of autonomous agents and, as such, has received much attention over the years. However, making optimal decisions and executing optimal actions typically requires a substantial effort on the part of an agent, and in some situations the agent might lack the necessary sensory, computational, or actuating resources to carry out the optimal policy. In such cases, the agent will have to do the best it can, given its architectural constraints. We distinguish between three ways in which an agent’s architecture can affect policy optimality. An agent might have limitations that impact its ability to formulate, operationalize (convert to internal representation), or execute an optimal policy. In this paper, we focus on agents facing the latter two types of limitations. We adopt the Markov decision problem framework in our search for optimal policies and show how gradations of increasingly constrained agent architectures create correspondingly more ...

• Type: Conference
• Year: 2003
• Where: ATAL
• Authors: Dmitri A. Dolgov, Edmund H. Durfee
• Keywords: ATAL 2003, Optimal Behavior, Optimal Decisions, Optimal Policy