Start small: Training controllable game level generators without training data by learning at multiple sizes

A level generator is a tool that generates game levels from noise. Training without dataset suffers feedback sparsity, since it unlikely to generate playable via random exploration. common solution shaped rewards, which guides the achieve subgoals towards playability, but they consume effort design and require game-specific domain knowledge. This paper proposes novel approach train generators datasets or rewards by learning at multiple sizes starting small up desired sizes. The denser negates need for rewards. Additionally, learn build various sizes, including were not trained for. We apply our recurrent auto-regressive generative flow networks (GFlowNets) controllable generation. also adapt diversity sampling be compatible with GFlowNets. results show create diverse Sokoban, Zelda, Danger Dave. When compared reinforcement better controllability competitive diversity, while being 9x faster training