Comparing the trade-off of believability and performance of abstract intelligent agents and humans playing Super Mario Bros.

Abstract intelligent agents within many genres of video games are crucial to a player’s enjoyment and immersion in their video game experience. There are two major aspects of abstract intelligent agents; performance and believability (Iskander, et al., 2012). Player believability refers to whether or not an observer of a character on screen believes that the character is being played by a human (Buche, et al., 2010; Togelius, et al., 2012).This research looks at how these two factors interact and whether or not there is a trade-off between the two, and where it is possible to produce an abstract intelligent agent that can be both highly believable and effective. We are exploring this issue using the Nintendo 1985 platform game Super Mario Bros as a benchmark. Showing video clips of agents built around different algorithms and human players of varying skill levels to 15 uninformed observers (Togelius, et al., 2012). These will provide insight into whether they thought the clip contained a human or agent player and why they thought so. Along with performance, the benchmarking of agents and players give is taken into account. The poster presentation will present preliminary results, which suggest that there is a compromise between believability and performance. At low level of performance the agent would have higher levels of believability when compared to a novice player. Although as the performance rises the less believable the agent becomes. This is due to the level of precision and effectiveness being too perceivably high for human to achieve. Further, we will present the traits that affect the relationship between believability and performance. These include decision-making speed, reaction time, accuracy, optimal and non-optimal approaches amongst others. Our work aims to initiate discussions on the interplay of the impact, the usage and design of intelligent agents for video games in an attempt to lead to believable agents that still fulfil their performance requirements