Stoix
Stoix copied to clipboard
🏛️A research-friendly codebase for fast experimentation of single-agent reinforcement learning in JAX • End-to-End JAX RL
## What? Added minimal support to stochastic muzero by issue #77. ## Why? To be able to train stochastic environments like 2048, poker, ... ## How? Added Afterstate and Encoder...
Add stochastic muzero implementation - [paper](https://openreview.net/pdf?id=X6D9bAHhBQ1) and the [pseudocode](https://gist.github.com/Mononofu/7548d8aa4bf94e12bc7eb7662fd60b56) With this improved version of muzero the stoic could be able to train stochastic environments like the 2048 game and poker...
## What? Add discrete SAC
### Feature Implement correct set up for multi-host systems as well as the current multi-device support. ### Proposal This involves using local devices and setting seeds appropriately using process ids.
### Feature [Revisiting Recurrent Reinforcement Learning with Memory Monoids](https://arxiv.org/abs/2402.09900) provides a method to combine recurrent models with standard, nonrecurrent RL losses. This should provide support for S5, LRU, FFM, Linear...
Just leaving this here as a reminder to change it properly when I have time.
### Problem In order to run a model, we need to specify the exact `.py` of the system model i.e `python stoix/systems/ff_ppo.py`. ### Solution It would be much easier if...
### Describe the bug Hello! When making the Dockerfile, I get the error `Cannot import name 'linear_util' from 'jax'` when running examples. This seems to be due to the incompatibility...