DQN Tackling the Game of Candy Crush Friends Saga : A Reinforcement Learning Approach

Abstract: This degree project presents a reinforcement learning (RL) approach called deep Q-network (DQN) for learning how to play the game Candy Crush Friends Saga (CCFS). The DQN algorithm is implemented together with three extensions, which in 2015 resulted in a new state-of-the-art on the Atari 2600 domain. This thesis shows that DQN in combination with the three extensions is an appropriate method for learning how to play CCFS. The influence that each of the extensions has on the performance is investigated separately, and reasoning for why or why not these extensions make sense in this environment is provided. CCFS is a stochastic game environment with many new features per level. This leads to a challenge when designing the reward function. This thesis investigates the impact of three different reward functions and reflects over why a certain type of design is more relevant. The results presented show that the DQN approach is able to learn a policy that increases its performance compared to that of random game-play. However, at this stage the performance is not yet reaching that of human game-play, but with further research we believe that it is possible.