An architecture for bounded rationality
Stuart Russell · ACM SIGART Bulletin · 1991
This paper is specifically concerned with the topic oflimited rationality--- the generation of utility-maximizing behaviour in a non-static environment by a system with finite computational resources. The design of the RALPH (Rational Agent with Limited Performance Hardware) architecture aims to tie together, in a unified theoretical framework and a single system design, three strands of research relevant to this problem: the study ofmetareasoningprocedures, that is, decision procedures that select computations to direct the course of another decision procedure; thecompilationof decision processes that explicitly maximize utility into efficiently-executable policies and goals, and their integration into the decision procedure; and the generation ofplanningbehaviour, directed towards an explicitly-represented goal, in the context of decision theory. Decision procedures are implemented by four distinctexecution architectures,running in parallel, each characterized by the types of knowledge employed. Only fragments of the architecture are currently implemented, but some progress has been made on the individual topics.