Applying Hindsight Experience Replay to Procedural Level Generation
Evan Kusuma Susanto, Handayani Tjandrasa · 2021
Designing a video game level requires a precise balance in difficulty adjustment as a level that is too simple will cause players to lose interest quickly. On the other hand, making a level too complicated will frustrate the players, making them abandon the game. We propose a new method to make a level generator that can learn how to design a game level by itself. Our proposed method can be used for different games with only minimal adjustments. We improve the previously proposed method by making our generator able to design a level that satisfies every user's criterion. We do this by combining Procedural Content Generation via Reinforcement Learning with the Hindsight Experience Replay method. We use our model to generate levels from 4 different games and compare the success rate with a random agent. Our model achieves more than 90% success rate for almost every scenario and performs much better when compared to a random agent.