Implement DeepRL agents

[epignatelli]

Agents are entities with a sample_action and update method, in potence. We exclude from the list exploration strategies and curricula.

Implement means either to produce new code from the paper directly, or to port an implementation from elsewhere, should that implementation be modular enough.

• Vanilla value learning

  • [x] Vanilla DQN
  • [x] SAC
  • [x] SAC Discrete
  • [ ] Categorical DQN
  • [ ] Double DQN
  • [ ] Dueling DQN
  • [ ] N-step DQN
  • [ ] Noisy DQN
  • [ ] Rainbow
  • [ ] DQN with Prioritised Experience Replay
  • [ ] Agent57
  • [ ] Expected Eligibility Traces

• Vanilla Actor critic

  • [x] A2C
  • [ ] DDPG
  • [ ] ACER
  • [ ] PPO
  • [ ] TD3
  • [ ] ACKTR

• Model-based

  • [ ] DreamerV1
  • [ ] DreamerV2
  • [ ] DreamerV3
  • [ ] AlphaZero
  • [ ] MuZero
  • [ ] Forward-Backward RL

• Credit assignment

  • [ ] RUDDER
  • [ ] Temporal Value Transport

• Hindsight methods

  • [ ] Hindsight Credit Assignment
  • [ ] Hindsight Policy Gradients
  • [ ] Hindsight Experience Replay
  • [ ] Upside-Down RL
  • [ ] Policy Gradients incorporating the future
  • [ ]

• Sequence modelling

  • [ ] Decision Transformers
  • [ ] Online Decision Transformers
  • [ ] Trajectory Transformer
  • [ ] UniMASK

• Distributed

  • [ ] R2D2
  • [ ] IMPALA
  • [ ] Seed RL

• Meta RL

  • [ ] RL2
  • [ ] Learned Policy Gradient
  • [ ] Algorithm Distillation

[epignatelli]

We should organise these by priority.