Simulating Human Game Play for Level Difficulty Estimation with Convolutional Neural Networks

Sammanfattning: This thesis presents an approach to predict the difficulty of levels in a game by simulating game play following a policy learned from human game play. Using state-action pairs tracked from players of the game Candy Crush Saga, we train a Convolutional Neural Network to predict an action given a game state. The trained model then acts as a policy.Our goal is to predict the success rate (SR) of players, from the SR obtained by simulating game play. Previous state-ofthe-art was using Monte Carlo tree search (MCTS) or handcrafted heuristics for game play simulation. We benchmark our suggested approach against one using MCTS. The hypothesis is that, using our suggested approach, predicting the players’ SR from the SR obtained through the simulation, leads to better estimations of the players’ SR.Our results show that we could not only significantly improve the predictions of the players’ SR, but also decrease the time for game play simulation by at least 50 times.