blog.openai.com

Meta-Learning for Wrestling

View CodeRead Paper We show that for the task of simulated robot wrestling, a meta-learning agent can learn to quickly defeat a stronger non-meta-learning agent, and also show that the meta-learning agent can adapt to physical malfunction. A simulated robot ant (red) uses its meta-learning policy to learn to beat a stronger creature with more legs (green). We've extended the Model-Agnostic Meta-Learning (MAML) algorithm by basing its objective function on optimizing against pairs of environments, rather than single ones as in stock MAML. MAML...

blog.openai.com

OpenAI and Microsoft

We‘re working with Microsoft to start running most of our large-scale experiments on Azure. This will make Azure the primary cloud platform that OpenAI is using for deep learning and AI, and will let us conduct more research and share the results with the world. One of the most important factors for accelerating our progress is accessing more and faster computers; this is particularly true for emerging AI technologies like reinforcement learning and generative models. Azure has impressed us by building hardware configurations optimized for...

blog.openai.com

Special Projects

Impactful scientific work requires working on the right problems — problems which are not just interesting, but whose solutions matter. In this post, we list several problem areas likely to be important both for advancing AI and for its long-run impact on society. We see these problems as having either very broad implications, or addressing important emerging consequences of AI development. If you are a strong machine learning expert and wish to start an effort on one of these problems at OpenAI, please submit an application. Detect if...