Reinforcement Learning (RL)
32 articles about Reinforcement Learning (RL)
OpenAI strengthens ChatGPT Atlas against prompt injection attacks
OpenAI is using automated RL-based red teaming to continuously find and patch prompt-injection exploits, hardening the ChatGPT Atlas browser agent as it becomes more agentic.
Introducing OpenAI o1, a large language model trained for complex reasoning
OpenAI o1 is a new reinforcement-learning trained LLM that improves complex reasoning by thinking before answering.
Procgen benchmark: 16 procedurally generated environments for evaluating reinforcement learning
Procgen Benchmark is a set of 16 easy-to-use procedurally generated environments to measure how fast reinforcement learning agents learn generalizable skills.
Neural MMO: a multiagent game environment for reinforcement learning
Neural MMO is a persistent, open-ended multiagent game environment for reinforcement learning that supports many evolving agents to improve exploration and competence.
Introducing Spinning Up: an educational resource for deep reinforcement learning
Spinning Up in Deep RL is a free educational resource with clear code examples, exercises, docs, and tutorials to help anyone learn deep reinforcement learning.
Agent achieves high score on Montezuma’s Revenge from a single demonstration
An agent learns Montezuma’s Revenge from one human demo, reaching 74,500 by training from selected demo states with PPO.
Evolved policy gradients: a metalearning method for fast adaptation to new tasks
Evolved Policy Gradients is an experimental meta-learning method that evolves an agent’s loss function to train faster and generalize to new tasks beyond its training setup.
SIMA 2: an AI agent that plays, reasons, and learns in virtual 3D worlds
SIMA 2 is a Gemini-powered AI agent that can reason, learn, and act with you in interactive 3D virtual worlds.
Training language models for summarization using human feedback
Using human feedback and reinforcement learning, we trained language models to produce better summaries.
Safety Gym: environments and tools for safe reinforcement learning
Safety Gym is a set of environments and tools to measure how well reinforcement learning agents learn while following safety constraints.
Review of the first Spinning Up in Deep RL workshop at OpenAI
A brief review of OpenAI’s first Spinning Up in Deep RL workshop held on February 2 as part of its new education initiative.
Reinforcement learning agents explore environments using prediction-based rewards
Random Network Distillation uses prediction-based curiosity rewards to drive exploration in reinforcement learning, achieving above-average human performance on Montezuma’s Revenge.
OpenAI Five neural networks begin defeating amateur teams in Dota 2
OpenAI Five, a team of five neural networks, is starting to beat amateur human teams at Dota 2.
Transfer learning contest evaluates reinforcement learning generalization
A transfer learning contest testing how well reinforcement learning algorithms generalize from past experience.
Addendum to o3 and o4-mini system card: codex coding agent
Codex is a cloud coding agent powered by codex-1 (an o3 variant) trained with reinforcement learning on real coding tasks to follow instructions, match human coding style, and run tests until they pass.
OpenAI co-organizes NeurIPS 2020 competitions on Procgen and MineRL
OpenAI and partners are co-organizing two NeurIPS 2020 AI competitions using the Procgen Benchmark and MineRL.
Neural networks enable a robot hand to solve the Rubik’s Cube
Neural networks trained in simulation use reinforcement learning and domain randomization to control a robot hand that solves a Rubik’s Cube and adapts to unexpected real-world disturbances.
Quantifying generalization in reinforcement learning using the CoinRun environment
CoinRun is a new reinforcement learning environment that measures how well agents generalize to new levels, balancing simplicity with a real transfer challenge.
Human-like robot hand trained to manipulate objects with advanced dexterity
A human-like robot hand has been trained to manipulate physical objects with unprecedented dexterity.
Full release of Gym Retro platform with over 1,000 games for reinforcement learning
Gym Retro is now fully released, expanding reinforcement-learning game research to 1,000+ games and providing a tool to add new games.
Showing page 1 of 2