3 articles
OpenAI publishes research on nonlinear computation in deep linear networks. The study explores how linear architectures can perform complex calculations despite the absence of traditional nonlinear activation functions.
OpenAI releases LOLA (Learning with Opponent-Learning Awareness), an algorithm that models other agents' learning and discovers collaborative strategies like tit-for-tat in the iterated prisoner's dilemma.
OpenAI introduces a reinforcement learning method where agents model opponent learning to improve strategy. Tested in multi-agent environments, this approach enables models to adapt behavior by anticipating adversary changes.