Tag

#game-playing

Every item tagged game-playing, newest first.

2 items

Dota 2 with large scale deep reinforcement learning

OpenAI trained a Dota 2 bot using large-scale deep reinforcement learning, achieving a win rate of 60% against human players in 1v1 matches. The bot learned from self-play with 180,000 hours of experience, equivalent to 20 years of continuous play. This demonstrates the potential of deep learning for complex tasks. You can apply similar techniques to other challenging problems.

Key takeaways

60% win rate against human players in 1v1 Dota 2 matches.
Bot learned from 180,000 hours of self-play experience.
Deep reinforcement learning applied to a complex game.

OOpenAI#deep-reinforcement-learning #game-playing #self-play

researchJul 4

Learning Montezuma’s Revenge from a single demonstration

Researchers trained an agent to achieve a high score of 74,500 on Montezuma's Revenge from a single human demonstration. The agent used a sequence of games from carefully chosen states and learned via PPO, a reinforcement learning algorithm. This result surpasses any previously published score. Builders can explore applying similar techniques to train agents for other complex tasks.

Key takeaways

Achieved 74,500 score on Montezuma's Revenge from one demo.
Uses PPO reinforcement learning algorithm.
Surpasses previous best published results.

OOpenAI#reinforcement-learning #single-demonstration #game-playing