• No results found

Experienced player-like behaviour identification

2.4 Impact of AI on the Player Experience

2.4.3 Experienced player-like behaviour identification

Evidence suggests that a judge's experience with a game being used in believability testing has an effect on the judge's reliability. For veterans of certain games, noticing the difference between opponents is usually more accurate (Livingstone, 2006). Research also suggests that experienced players have a better understanding of the objectives of the game which can result in the identification of specific, player-like behaviour. As an example, highly skilled and experienced players are found to traverse the game environment quite methodically, collecting items to strengthen their character to give them a statistical advantage during future enemy encounters (Gorman & Thurau, 2006). Because of this, it is worth exploring the actions and behaviours of experienced players of FPS games as well as the mechanics and ideologies which can help represent this player-like behaviour.

Early research by Laird et al. involved the study of skill and humanness levels in a custom developed bot 'Soarbot' (Laird et al., 2000). Specifically it looked at how efficient and humanlike the bot's aiming is. Importantly, Laird also mentions that actual human players who are playing the game may not actually be the best judges, being too occupied with the game and thus unable to devote much time to observing players closely. While this is an important distinction, it is also necessary to point out that using external judges inexperienced with the game is as equally as erroneous, evidenced by results in the first BotPrize competition (Hingston, 2009). To address this, human evaluators were given recordings of what the bot sees so they could compare it to how the humans behaved in the same situation. During the testing, it was found that decision times hypothesized for human players were observed to be the most human-like, however aiming skill was not a good indication of human-like behaviour.

37

Continuing from this, Laird later assessed the aggressiveness, aiming skill, decision time and tactical knowledge (in a Turing test fashion) of bots and concluded that variations in decision time show changes in ratings of humanness. The best performance appeared to come from a decision time similar to that hypothesized for humans (Laird, 2001). One of several aspects they tested for was the level of perceived skill and humanness of the bot with a set decision time. At low decision time intervals, the bot‟s skill was seen to be impressive. However, at the same time this was not seen to be very player-like in execution. A fast reaction time, while humanly possible in certain circumstances, does not belie a player-like manner of execution. In human players, the fast rotation and accuracy demonstrated by bots is both unreliable and almost impossible to perform (Conroy & Wyeth, 2010; Rayner, 2007). In order to achieve a human level of expert skill, additional improvements were needed in other skill areas because the decision time for their fastest example was superhuman. Laird also explained that they needed to improve the time modelling of sensing and motor actions as well as improve other aspects of the bot‟s skill (Laird, 2001). The algorithms and techniques developed by Laird controlling the bot's aim are similar to those used in games today.

Just as experienced players are more adept at identifying AI opponents, so does their appreciation for the decisions made by the AI. The work by Johnson and Gardner goes into detail regarding aspects of the media equation, but specifically describes a user study involving teams of either humans or both human and computer players (Johnson & Gardner, 2005). The most significant finding was that the more experienced a human player was, the less inclined they were to listen to recommendations by the AI. Additionally, they were more inclined to rate the usefulness of the computer lower and generally perceived themselves as less similar to the computer players. This is similar to findings made by Krach et al. (2009), who identified that engagement levels drop considerably when interfacing with an alleged AI opponent (Krach et al., 2009). Krach's research reveals interesting results about how people behave and feel when they think their opponent is human. Greater Theory of Mind (ToM) brain activity was present when players played a game of rock-paper-scissors against an alleged human player. However there were differences in brain activity depending on the gender of the human player. In particular, women

38

were viewed to not be as engaged as men when playing against an alleged computer opponent, whereas men compensated for weaker ToM abilities with increased effort.

Finally, the research by Hingston and Soni focuses on a qualitative approach to garnering how experienced players perceive bot actions. Hingston's team designed bots for a commercial video game that uses a neural network to select actions. They recorded data from a human player and used it to train the network. They then tested the resulting bots by having human players compete against them, and asking them about their impressions of their opponents (Soni & Hingston, 2008). The results were successful with the human testers finding the trained AI to be more human-like in nature. Their results showed that in the context of their study, human players consistently found the bots to be more player-like, less predictable, more re-playable, and more challenging than the provided, hand-coded bot (Soni & Hingston, 2008). While Hingston's research is primarily focused on identifying player-like behaviours, it is but one of many examples of how different AI technology can be used to create better bot AI opponents. The topic of FPS bot AI research and technology is therefore worth exploring.