Evaluating the Robustness of GAN-Based Inverse Reinforcement Learning Algorithms

We evaluate the robustness of reward functions learned with IRL, when transferred to similar tasks. We exceed state of the art results for one benchmark task and solve another one for the first time. Modifications are proposed that achieve faster and more stable training.

Detalles Bibliográficos
Autor: Heidecke, Johannes
Tipo de recurso: tesis de maestría
Fecha de publicación:2019
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/131625
Acceso en línea:https://hdl.handle.net/2117/131625
Access Level:acceso abierto
Palabra clave:Reinforcement learning
Algorithms
inverse reinforcement learning
IRL
reinforcement learning
RL
guided cost learning
GCL
adversarial inverse reinforcement learning
AIRL
soft actor critic
SAC
transfer learning
robustness
maximum entropy principle
maximum causal entropy principle
reward shaping
pre-training
metric
shaped reward loss
pendulum
lunar lander
Aprenentatge per reforç
Algorismes
Àrees temàtiques de la UPC::Informàtica
Descripción
Sumario:We evaluate the robustness of reward functions learned with IRL, when transferred to similar tasks. We exceed state of the art results for one benchmark task and solve another one for the first time. Modifications are proposed that achieve faster and more stable training.