Is ‘minimax’ an Effective Autocurriculum in JAX?

Original title: minimax: Efficient Baselines for Autocurricula in JAX

Authors: Minqi Jiang, Michael Dennis, Edward Grefenstette, Tim Rocktäschel

In the quest to train smarter decision-making AI, there’s a technique called Unsupervised Environment Design (UED). It helps agents adapt to new situations they’ve never seen before. But here’s the snag: the usual experiments take forever to run—weeks on end! That’s like trying to sprint through molasses. Enter minimax—a new hero in the form of a library. It’s built to turbocharge these experiments, making them fly on faster hardware. Using JAX, it revs up the training loop, slashing those weeks of waiting down to a fraction of the time. And it’s not just speed; it’s about creating a playground for trying out new ideas quickly. With minimax, researchers get a petri dish for experimentation, complete with tools to tinker with procedurally generated environments. Plus, it comes with solid starting points for these experiments, making them run over a hundred times faster than before. It’s like hitting fast-forward on AI training.

Original article: https://arxiv.org/abs/2311.12716