There are several areas in video games in which it is desirable to emulate human likeness of navigation. Non-player characters (NPC) and artificially intelligent (AI) opponents can become more believable and therefore more fun to play with or against when perceived as humanlike. Furthermore, as triple-A games grow larger, game testing becomes increasingly important and the ability to automatically test games becomes more valuable. Transfer learning (TL) can help speed up the training times of Deep Q-network (DQN) agents and thus lower the costs of developing game-testing- or NPC/enemy AI. There is also an interest in general video game AI that TL and DQNs can help begin to examine.

This study seeks to examine the transferability of properties that affect the human likeness of game playing DQN agents. It does so by presenting an experiment in which a group of experts evaluated human likeness across several agents trained with different hyperparameters for the same game. After identifying the most humanlike agents they were transferred to a second game for further training alongside a non-transferred baseline trained agent using the same hyperparameters. A second set of observations were made contrasting the human likeness of the transferred agents with the baseline to see how the transfer affected the human likeness of the agents.

From the thematic analysis four themes emerged: Intent, Personality, Intellect, and Movement (IPIM). These themes were found effective and exhaustive when evaluating perceived human likeness from field notes and can provide a good basis for further analysis of human likeness in game playing agents. Furthermore, the analysis mapped out how certain hyperparameters affect the human likeness of game playing AI agents interpreted through those themes.

Given the limitations of this study we are unable to conclusively answer the research questions. Our findings suggest that the shape of the reward function has an effect on the trained human likeness after transfer, but further studies are necessary to either confirm or reject these findings.