Constrained Inverse Reinforcement Learning: Challenges and Solutions for Real World Implementation

Kamgarpour, MaryamSchlaginhaufen, AndreasWood, Tony AlanRen, KaiChassagne, Pierre2023-08-242023-08-242023-08-242023-06-14https://infoscience.epfl.ch/handle/20.500.14299/200051This thesis explores the challenges and solutions linked to the implementa- tion of Constrained Inverse Reinforcement Learning (CIRL) for real world application. To this end we study two algorithms, one utilizes stochas- tic gradient descent ascent (SGDA-CIRL), while the other incorporates IQ- Learn, an advanced imitation learning algorithm. Findings reveal that the Q-CIRL algorithm shows potential and succeeds in recovering a reward for which the expert is optimal but fails to generalize to new transitions dynam- ics. Meanwhile, the SGDA-CIRL algorithm demonstrates fast convergence and results comparable to CIRL with known dynamics. Additionally, an open-source framework for CIRL is developed, providing a versatile plat- form for implementing and extending CIRL algorithms and reinforcement learning techniques.Inverse Reinforcement LearningConstrained Inverse Reinforcement Learning: Challenges and Solutions for Real World Implementationstudent work::master thesis