Implementation of DQN, n-step DQN and TreeQN. Tested on Cartpole and various Atari. Reproduces results in TreeQN and fixes a subtle bug in the authors' implementation Contains the code for an abandoned project. Important feature: Modular code for easy addition of custom losses (such state prediction loss, reward loss, etc). TODO: clean up some code, refactor models for readbility.
srinivr / rl Goto Github PK
View Code? Open in Web Editor NEWImplementation of DQN, n-step DQN and TreeQN