4 articles
OpenAI demonstrates mathematical equivalence between policy gradient methods and soft Q-learning in reinforcement learning. This theoretical finding unifies two major RL approaches and enables combining their respective advantages.
OpenAI publishes research on stochastic neural networks for hierarchical reinforcement learning. The method improves agents' ability to decompose complex tasks into sub-objectives.
OpenAI developed an unsupervised system that learns an excellent sentiment representation by being trained only to predict the next character in Amazon reviews.
OpenAI has developed a spam-detection AI system trained entirely in simulation and deployed on a physical robot. First application of its kind capable of operating in the real world.