Inverse Reinforcement Learning with Gaussian Processes

Please enable javascript in your browser.

Habibi, Beheshteh | 2022

210 Viewed

Type of Document: M.Sc. Thesis
Language: Farsi
Document No: 54972 (02)
University: Sharif University of Technology
Department: Mathematical Sciences
Advisor(s): Sharifi Tabar, Mohsen
Abstract:
Inverse reinforcement learning (IRL) is one of the machine learning frameworks based on learning from humans; That is, instead of producing a decision process maximizing a predefined reward function, seeks to find the reward function based on the observed behavior of an agent. The biggest motivation of IRL is that, usually, determining a reward function for a problem is very diﬀicult. We consider IRL in Markov decision processes; that is, the problem of extracting a reward function with the assumption of knowing the optimal behavior. IRL could be useful for apprenticeship learning to obtain skilled behavior, and for optimizing a reward function by a natural system. We first, determine a set of all reward functions for which the given policy is optimal and introduce algorithms in discrete-finite and continuous-infinite cases. Although most IRL algorithms represent reward as a linear combination of the set of features we use Gaussian processes as a nonlinear function to learn the reward. Then, we will compare different algorithms for solving this problem
Keywords:
Reward Function ; Optimal Policy ; Gaussian process ; Inverse Reinforcement Learning (IRL) ; Markov Decision Making

No TOC