Reinforcement Learning (RL)
32 articles about Reinforcement Learning (RL)
Release of eight simulated robotics environments and hindsight experience replay implementation
We’re releasing eight simulated robotics environments, a Hindsight Experience Replay baseline, and a set of robotics research requests to help train models that transfer to real robots.
OpenAI releases baselines implementations for ACKTR and A2C algorithms
OpenAI Baselines adds ACKTR and A2C, with A2C matching A3C performance and ACKTR offering better sample efficiency than TRPO/A2C with only slightly more compute.
Proximal policy optimization: a simple and effective reinforcement learning algorithm
Proximal Policy Optimization (PPO) is a simple, easy-to-tune reinforcement learning algorithm that matches or beats top methods and is OpenAI’s default due to strong performance.
Evolution strategies as a scalable alternative to reinforcement learning on modern benchmarks
Evolution strategies can match standard reinforcement learning on benchmarks like Atari and MuJoCo while avoiding many of RL’s practical drawbacks.
Hierarchical reinforcement learning algorithm discovers high-level actions for navigation tasks
A hierarchical reinforcement learning algorithm learns reusable high-level actions like walking and crawling to solve new navigation tasks much faster.
Training AI with occasional human feedback using RL-Teacher
RL-Teacher is an open-source tool that trains reinforcement learning agents using occasional human feedback instead of hand-crafted reward functions, especially when rewards are hard to define.
OpenAI releases baselines for DQN and its variants
OpenAI open-sourced its Baselines project, starting with DQN and three variants to match published reinforcement learning results.
Common issues caused by misspecified reward functions in reinforcement learning
This post explains how reinforcement learning can fail in unexpected ways when the reward function is misspecified.
Meta-learning agent adapts and outperforms in simulated robot wrestling
A meta-learning robot wrestling agent quickly learns to beat stronger opponents and adapts to physical malfunctions in simulation.
Improving reinforcement learning exploration with adaptive parameter noise
Adding adaptive parameter noise to reinforcement learning algorithms improves exploration and often boosts performance with minimal downside.
Roboschool: open-source robot simulation software integrated with OpenAI Gym
Roboschool is an open-source robot simulation software release integrated with OpenAI Gym.
OpenAI releases public beta of Gym toolkit for reinforcement learning
OpenAI Gym Beta is a public toolkit with many environments and a results site to build, test, and compare reinforcement learning algorithms.
Showing page 2 of 2