Evolved Policy Gradients

OpenAI has introduced Evolved Policy Gradients (EPG), a metalearning technique that uses evolutionary algorithms to automatically discover loss functions for training reinforcement learning agents. Unlike traditional fixed loss functions, EPG evolves the loss function over generations, enabling agents to learn more efficiently and generalize to novel tasks at test time. In experiments, agents trained with EPG successfully completed navigation tasks in configurations they had never encountered during training, such as reaching an object placed on the opposite side of a room. This approach falls under metalearning, where the goal is to learn how to learn. For developers building AI workflows, EPG suggests a path toward more adaptive agents that require less manual tuning and can handle a wider variety of tasks without retraining. The method automates a key component of training, potentially reducing the engineering effort needed to deploy reinforcement learning in real-world applications.