All Life is Problem Creation: Learning to Generate Environments that Maximize Performance Gain

Intelligent agents can achieve mastery not just by learning on well-defined problems, but also by creating their own experiences that maximise learning. While current methods for automatic curriculum generation often rely on heuristics such as task novelty or difficulty, these proxies are often misaligned with the ultimate task. An agent can be endlessly captivated by novel-but-unlearnable environments or stymied by difficult-but-irrelevant challenges. We propose a framework where a generative Proposer agent learns to create environments that explicitly maximise Solver agent's performance gain on a target task. To make the curriculum adaptive, the Proposer is conditioned on the Solver's policy, obtained by probing its behaviour on a small set of diagnostic environments. This conditioning mechanism enables the Proposer to generate a sequence of training environments, targeting the Solver's evolving weaknesses. We validate our approach in maze environments, where our method learns to generate a curriculum of environments that are distinct from the target task distribution. Our experiments demonstrate that this approach accelerates the Solver's learning on both in-distribution and out-of-distribution tasks compared to training directly on the target distribution.