Estimating Player Completion Rate in Mobile Puzzle Games Using Reinforcement Learning

In this work we investigate whether it is plausibleto use the performance of a reinforcement learning (RL) agentto estimate the difficulty measured as the player completion rateof different levels in the mobile puzzle game Lily’s Garden.For this purpose we train an RL agent and measure thenumber of moves required to complete a level. This is thencompared to the level completion rate of a large sample of realplayers.We find that the strongest predictor of player completion ratefor a level is the number of moves taken to complete a level of the∼5% best runs of the agent on a given level. A very interestingobservation is that, while in absolute terms, the agent is unable toreach human-level performance across all levels, the differencesin terms of behaviour between levels are highly correlated to thedifferences in human behaviour. Thus, despite performing sub-par, it is still possible to use the performance of the agent toestimate, and perhaps further model, player metrics