Advanced Ray RLlib: New Features and New Possibilities
summary
RLlib now ships PyTorch for all algorithms plus native TensorFlow 2 eager support and Windows wheels, and adds model-based methods Dreamer and MB-MPO. Dreamer reproduces DeepMind results, hitting target rewards after ~5 million real environment steps versus ~100 million for model-free baselines (≈20× sample efficiency). A new exploration API enables Thompson sampling and UCB contextual bandits (AWS contrib), plug-in curiosity via intrinsic prediction-error rewards for sparse mazes, a Unity adapter, external client/server mode, and trajectory/attention work aimed at stability and speed.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 0:31Reinforcement Learning Cycle
- 2:34New Features
- 4:36Model Based RL
- 8:28Exploration API
- 11:10Curiosity API
- 15:10Complex Environments
related talks
11mTorchRL: The Reinforcement Learning and Control library for PyTorch
18mLightning Talk: TorchRL - RLHF Support
23mCost-Effective RL Post-Training with Ray and AWS Trainium
28mScaling Multimodal Data and Reinforcement Learning with Ray
19mA 24x Speedup for Reinforcement Learning with RLlib + Ray
30mTransforming Grab’s AI Platform with Ray
29mMiles: Enterprise-Facing Agentic RL Framework
26mChatGPT and Minecraft's diamonds playing agents: Reinforcement Learning in a nutshell with sheeprl