In our MIG 2013 paper “Evaluating perceived trust from procedurally animated gaze”, we investigated how well a simple gaze model conveyed how much a character appeared to trust the viewer. If the character trusts the player a lot, they spend the more time looking at the face; otherwise, they spend more time looking away. Check out the video to see a demo of the gaze model and the experiment platform we built in Unity to make it.
In the above work, we find that even with very fast animation clips, people naturally inferred attitudes from the character solely based on the proportion of time the character spent looking at the player. Specifically, gaze conveyed trust, interest, admiration, and friendliness to varying degrees.
Above, * represents significance at the 0.05, 0.01, and 0.001 probabilities, where the null hypothesis corresponds to a regression slope of 0 (e.g. a horizontal line). In other words, the steeper the slope which fits viewer rankings to time looking directly at the player, the more significant the result.
Given how straight-forward gaze is to implement, it’s clear that it’s an easy way to add personality to characters. Additionally, in the paper, we show that it’s also straight-forward to tie gaze to character attitudes which might vary dynamically while you play a game — all without needing additional voice acting, motion capture, scripting, and dialog. For example, suppose a player had a high reputation score with a character’s group, or that the player has spoken with the player multiple times, the character’s body language and voice tone could automatically reflect that. We also show how the gaze can vary probabilistically to look more natural, while still maintaining a desired proportion of time looking in a desired direction.
Player modeling can be used to train NPCs and bots, to dynamically customize the gameplay (for example, an enemy’s strategies could change based on play style), and to aid testing and level design. In our 2013 AIIDE paper, we proposed a simple probabilistic method for modeling players that could be used to bias players towards certain behaviors. The underlying assumption is that players tend to act certain ways based on what is available in their environment. Thus, if we know the relationship between player behavior and environment, we can tweak the environment to encourage people to behave in certain ways.
In other words, we can model what players do where and then use this information to nudge player behaviors in desired ways. There may be applications for this beyond testing and data collection — video games are unique in that we have absolute control over the environment we present to players. For example, could we better understand what environmental/game incentives either encourage (or discourage) PVP? What differences are there between free-to-play players and subscribers? What do players tend to do at max level?
For the paper, we specifically looked at a straight-forward application of this idea for collecting player metrics. Such an approach could reduce the number of games playtesters need to run because it would allow them to focus on collecting data only for the metrics which need it most. For our proof of concept, we implemented several dynamically configurable environments in Second Life and collected several very simple behavior metrics: the distances between people standing in either narrow or wide spaces; the timing of lane crossings for slow and fast traffic; and the choice of whether to use a health kit based on health level.
Below are two screenshots from two of our experiment setups (top: a space environment; bottom: an office environment), in which players race around to collect tiles for prizes.
Using our player model, we formulated the question of which game to run next as an optimization problem (as an MDP), would run more games, and finally update our player model using the results. Even without running the optimization, one can look at the statistics to see what behaviors occur most frequently in which environments. Even in our straight-forward setup, our assumptions about what players would do were often wrong! We also showed that our optimization-based scheduler did reduce the number of games needed to run, when compared to a schedule which played all scenarios equally. However, there are caveats and limitations to the approach which are worth reading about in the paper.
Even for testing, we didn’t get a chance to explore these ideas further, but I always envisioned it having potential in large, open-world multiplayer environments where it’s particularly difficult to work out every glitch or even understand apriori how players will interact with each other and the game (although companies do an amazing job). For example, could this approach help debug aspects of the environment that lead to trapping players in walls (e.g. do environments where players become trapped have certain shared characteristics? or are the aspects of the character (such as speed at impact) cause the character to become stuck)? Could this approach help debug navigation problems for companion characters, who may block and trap the player in certain areas? After all, once a problem is understood well-enough to be reproducible, it’s often straight-forward to fix.