Bootstrapped DQN

This repository maintains the code for Bootstrapped DQN (Osband 2016).

Experiments are being run to check the best suited version (DDQN/DQN, with/without gradient normalisation, number of heads) and the plots will be put up here soon :)

The current implementation uses DQN with Bootstrap architecture.

Preliminary Graphs

Breakout 2 Heads

The current implementation involves using 2 heads without gradient normalisation for 100 epochs.

Breakout 10 heads (with and w/o gradient normalization)

This is the first result with 10 heads with BootstrappedDQN.