Jigsaw2 Solver Demo. My theory is that world model RL which relies on aggregate situational representation, such as DreamerV3, cannot manage combinatorial complexity. Specifically, it cannot learn to distinguish details — such as jigsaw curves — and their detail specific affordances. This environment provides an almost purely perceptual task, making the action space small (consider next piece, rotate considered piece, attempt match, toggle view) and the dynamics simple. It also provides the agent with lots of learning aids such as a piece focused fovea, highlighting match edges, and match attempt animation that translate attempted match edges towards each other before announcing success or failure. I suspect Dreamer will succeed in a finite condition with repeating puzzle edges where it can memorize edge matches in training. But despite learning aids and simple dynamics, I expect Dreamer will fail in the infinite condition with always novel edges. This simple hand engineered solver demonstrates that the task is readily solvable in the infinite condition with the same data and actions available to Dreamer.

Jigsaw Solver Demo. On left: overview of entire Jigsaw canvas with fovea outlined in orange, and green cross for action cursor. On right: fovea contents. The solver first scans the canvas to build its own complete representation, calculates matches, and then piecewise solves the puzzle.