implement n-steps in buffer class, to be reused by different agents

lucasosouza / fasterRL

Reinforcement Learning library based on pytorch. Designed for research and experiments. Platform-agnostic (supports openai, marlo, more to be added)

MIT License

6 stars 1 forks source link

implement n-steps in buffer class, to be reused by different agents #2

Open lucasosouza opened 5 years ago

lucasosouza commented 5 years ago

also implement importance sampling to make it useful in off-policy methods without making it online which would not work well in experience buffer where experiences can be really old