Rainbow

[seann999]

Some changes to enable Rainbow

• Merges Adds N-step learning for DQN-based agents. #317 for n-step learning
• Adds prioritized replay buffer weights for CDQN using KL-based loss
• Adds Double CDQN agent
• Adds dueling architecture for CDQN
• Adds noisy net option for CDQN ALE example
  Thus, Rainbow can be used in CDQN example with e.g.
  python3 examples/ale/train_categorical_dqn_ale.py --gpu 0 --arch dueling --noisy-net-sigma 0.5 --agent DoubleCDQN --num-step-return 3 --env BattleZoneNoFrameskip-v4 --prioritized

Currently running experiment in BattleZone game to compare with CDQN baseline