A Framework for Bounded Rationality and Limited Resource Sensing for Optimized Decision Making in Artificial Intelligence

Abstract

Bounded rationality acknowledges the constraints on an agent's cognitive resources and information availability, diverging from the notion of perfect rationality. Bounded Rationality in Artificial Intelligence (BRAI) involves stateful resources, where the agent's actions can alter the state of the resources, leading to observational bias and affecting the quality of information collected. We investigated how Artificial Intelligence (AI) agents could make optimal decisions despite various constraints, introducing a new perspective on bounded rationality specifically tailored for AI, the so called BRAI. This research addresses the concept of BRAI, focusing on the Limited Resource Sensing Problem (LRSP) and its impact on decision-making processes. We formalized BRAI using a Hidden Markov Model (HMM) framework, which accommodates the stochastic nature of stateful resource behavior. This methodology leverages reinforcement learning to develop a sensory controller that optimizes the agent's sensing activities, ensuring a balance between the benefits of extensive sensing and the drawbacks of resource state changes. The developed approach enhanced the agent's decision-making capabilities by refining the quality of information used while mitigating the potential for biased observations due to bounded rationality. This study introduced a novel framework for understanding and addressing bounded rationality in AI, providing a pathway for developing AI agents capable of making more informed and less biased decisions under resource constraints.