yhl04/agent57

An unofficial implementation of Agent57 for Atari from DeepMind

Agent 57

An unofficial implementation of Agent57 for Atari from DeepMind

Breakout Test Run

Features

• Double Q Network
• Dueling Network Architecture
• Distributed Architecture (Multi Actor)
• Update new recurrent states after training step
• LSTM Recurrence
• Episodic Novelty Module
• Lifelong Novelty Module
• Separate Nets for extrinsic and intrinsic reward
• Retrace (Replaced N-Step)
• Reward value rescaling
• Batched Inference (SEEDRL)
• Adaptive Exploration with meta-controller
• Prioritized Experience Replay
• Find bugs and test for correctness

Relevant Papers

Recurrent Experience Replay in Distributed Reinforcement Learning (R2D2):

https://openreview.net/pdf?id=r1lyTjAqYX

Never Give Up: Learning Directed Exploration Strategies (NGU):

https://arxiv.org/pdf/2002.06038.pdf

Agent57: Outperforming the Atari Human Benchmark (Agent57):

https://arxiv.org/pdf/2003.13350.pdf

Helpful Repositories