3 ms·
The paper, “Hybrid Reward Architecture for Reinforcement Learning”[0], describes an evolution on DeepMind's Deep Q-Network (DQN) design[1] that was able to play
by espadrine 9y ago
The paper, “Hybrid Reward Architecture for Reinforcement Learning”[0], describes an evolution on DeepMind's Deep Q-Network (DQN) design[1] that was able to play many Atari games, presented in early 2015.
The HRA design requires more preprocessing (DQN worked on the raw pixels and on score alone), but the charts mapping how much faster HRA learns and how much stronger it plays is convincing.
[0]: https://arxiv.org/pdf/1706.04208.pdf https://arxiv.org/pdf/1706.04208.pdf
[1]: https://deepmind.com/research/dqn/ https://deepmind.com/research/dqn/