DQN is augmented with a reward function (EvoReward) that evolves (each episode currently) using a basic genetic algorithm.

EvoReward takes in the current environment state and agent action to provide a good direction for the agent, even in the absence of external rewards. Experiments with Q-learning and Deep Q-learning agents augmented with EvoReward show that it significantly improves performance in both dense and sparsely rewarded environments and that it has the potential to be practically used.

python 3.5
Keras
gym
gym_minigrid
numpy
matplotlib

gym_minigrid is a minimal gym environment with sparse rewards.

Runs DQN+EvoReward by default with a population of 10 on MiniGrid-Empty-6x6-v0 environment. Use '--dqn_type 0' argument to use a normal DQN to compare. Use '--help' to see more options.

• evoDQN.py: DQN implementation supporting EvoReward function
• evoReward.py: Evolving reward function class
• multiprocess.py: For multiprocessing each individual in the population. This is test code. Does not work with EvoReward currently.
• parallel_trials.py: Aggregate results from multiple trials of training.
• scripts/run_grid_trials: Script to run parallel_trials.py in multiple processes to get results faster.

Formal algorithm and results can be found in docs