A PyTorch implementation of Protagonist Antagonist Induced Regret Environment Design
This codebase provides a PyTorch implementation of Protagonist Antagonist Induced Regret Environment Design (PAIRED), which was first introduced in “Emergent Complexity and Zero-Shot Transfer via Unsupervised Environment Design” (Dennis et al, 2020). This implementation comes integrated with custom adversarial maze environments based on MiniGrid environment (Chevalier-Boisvert et al, 2018), as used in Dennis et al, 2020.
Unsupervised environment design (UED) methods propose a curriculum of tasks or environment instances (levels) that aims to foster more sample efficient learning and robust policies. PAIRED performs unsupervised environment design (UED) using a three-player game among two student agents—the protagonist and antagonist—and an adversary. The antagonist is allied with the adversary, which proposes new environment instances (or levels) aiming to maximize the regret of the protagonist, estimated as the difference in