DN
Dhruv Nigam
@dhruvn
Dota 2 with Large Scale Deep Reinforcement Learning
Submitted Apr 17, 2024
On April 13th, 2019, OpenAI Five became the first AI system to defeat the world champions at an esports game. The game of Dota 2 presents novel challenges for AI systems such as long time horizons, imperfect information, and complex, continuous state-action spaces, all challenges that will become increasingly central to more capable AI systems.
Reinforcement learning has been a branch of machine learning with a lot of potential that has not materialized. It has proven instrumental in games where the environment is well-defined and can be simulated without cost. The real world does not offer these benefits. However, RLHF was a key component in making ChatGPT a success.
It is important to understand RL fundamentals to create systems that can continually improve from online feedback. I will cover through this paper -
- the philosophy or RL
- the challenge and importance of rewarding engineering
- The role of distributed training and Ray in making RL viable
{{ gettext('Login to leave a comment') }}
{{ gettext('Post a comment…') }}{{ errorMsg }}
{{ gettext('No comments posted yet') }}