A few of the key advantages of the Rainbow agent are summarized here for your quick reference:
- Combines several notable extensions to Q-learning developed over the past several years
- Achieves state-of-the art results in the Atari benchmarks
- n-step targets with a suitably tuned value for n often leads to faster learning
- Unlike other DQN variants, the Rainbow agent can start learning with 40% less frames collected in the experience replay memory
- Matches the best performance of DQN in under 10 hours (7 million frames) on a single-GPU machine
The Rainbow algorithm has become the most sought after agent for discrete control problems where the action space is small and discrete. It has been very successful ...