Bounded Rationality in Reinforcement Learning
: (Alternative Format Thesis)

  • Jan Malte Lichtenberg

Student thesis: Doctoral ThesisPhD

Abstract

The broad problem I address in this dissertation is the design of autonomous agents that can efficiently learn goal-directed behavior in sequential decision-making problems under uncertainty. I investigate how certain models of bounded rationality—simple decision-making models that take into consideration the limited cognitive abilities of biological and artificial minds---can inform reinforcement learning algorithms to produce more resource-efficient agents. In the two main parts of this dissertation I use different existing models of bounded rationality to address different resource limitations present in sequential decision-making problems. In the first part I introduce a boundedly rational function approximation architecture for reinforcement learning agents to reduce the amount of training data required to learn a useful behavioral policy. In the second part I investigate how Herbert A. Simon's satisficing strategy can be applied in sequential decision making problems to reduce the computational effort of the action-selection process.
Date of Award29 Mar 2023
Original languageEnglish
Awarding Institution
  • University of Bath
SupervisorÖzgür Şimşek (Supervisor) & Michael Tipping (Supervisor)

Keywords

  • reinforcement learning
  • bounded rationality
  • machine learning
  • artificial intelligence

Cite this

'