No announcement yet.

Data from multiplayer video games

  • Filter
  • Time
  • Show
Clear All
new posts

  • Data from multiplayer video games

    This thread is for discussing the potential of using data from multiplayer video games as a basis or part of a dataset for prosocial behavior.
    My initial thoughts on the matter:

    Quick argument
    Cooperating with video game companies could grant access to huge amounts of data of (mostly) anti-social behavior from the banning/punishment process of online multiplayer games. Furthermore, video game companies are likely to have valuable insights into what is considered ethical behavior. See e.g. this excerpt from a post by Riot Games:
    For example, report data shows the vast majority of players can’t stand racism, sexism and homophobia […]. In English speaking territories, “Your mom” can be silly or mean, but in Korea, “Your mom” is always deeply offensive.

    As machines become more capable, they will be autonomous members of mixed human-machine teams, playing part in leadership, decision-making, influencing team morale, or disciplining other team members. Beyond the interaction with team members, there are also scenarios where the behavior towards actors that are not part of the own team is relevant, e.g. when facing an opposing team in a competition. Video games provide a vast number of both "purely cooperative" and competitive scenarios, though I would expect competitive scenarios to be overrepresented in video games compared to practical real-world applications. Furthermore, competitive single-player games can provide examples of pro-social or anti-social behavior relating to "enemies".

    From the overview:
    We are trying to provide sufficient good-enough examples of pro-social behavior to teach machines to behave in a manner comparable to that of a well-raised young human child or friendly dog.
    The description is very broad and abstract and I currently have no mental model of what a data set that can teach pro-social behavior from scratch would look like. In video games, on the one hand, one can generally assume that all participants already have a good idea of what good behavior looks like, and accidental anti-social behavior is probably comparatively rare. On the other hand (for reasons I don't want to further speculate on), overt anti-social behavior is much more prevalent in video games than in real life scenarios and probably usually intentional, i.e. with the aim to aggravate other players. As such a dataset from a video game may be more applicable for a later stage in training, than for a basis.

    Every online multiplayer game of at least moderate size has some method to punish misbehaving players [citation needed]. Typical punishable offenses are cheating, insulting other players, offensive language, and intentionally helping the enemy team. Some of these behaviors can be universally recognized (e.g. offensive language) whereas others are very specific to the game's mechanics, rules, and the context (e.g. some forms of cheating may be explicitly allowed in certain game modes). The more specific the reason for punishment is documented, the easier it is to distinguish universally good and bad behavior from game-specific good or bad behavior.
    In addition to punishment, some games also have methods for rewarding players for pro-social behavior. One could also tentatively label behavior that did not lead to punishments as pro-social. For instance, if a player was never punished after X hours of play, it may be reasonable to label all of their collected data as "pro-social".

    Data format:
    • Many games have some method for chatting/messaging, and the pure text data could potentially also be integrated into other text databases, e.g. from online forums.
    • Having a record of the gameplay can provide 2D or 3D data on interaction between agents, though this data will be very game-specific and probably difficult to train transferable behavior on.
    • Using more "abstractable" information such as some scoring measure over time, could be an approach to make data from different games/scenarios comparable.

    • Games develop their own culture and standards for good and bad behavior, and only a subset will be universal.
    • As evidenced by the discrepancy between the prevalence of anti-social behavior in online video games and real life, behavior can be highly context-dependent, and it is not clear how applicable lessons learned from a dataset as outlined here are.

    Concrete Example: League of Legends
    League of Legends/Riot Games has access to literally millions of matches played, most of them between teams of five players. In addition to reports for bad behavior, there is also an "honor" system, rewarding players for good behavior.
    I myself haven't been active in online multiplayer games for a while now, and I don't really know many more details for the system in League of Legends. It would be great to hear from others who are more familiar with the report and banning systems from any video games to get more details on what sort of data might already be available and facilitate discussion of applicability and feasibility of using such data.

    • Massive amounts of data labelled by humans

    • Most labelled data anti-social
    • Transferability unclear
    • Access to data and privacy concerns could be difficult
    • Not clear whether "basic" enough

    Final random side note: Physical sports also have punishments for unsportsmanlike behavior and data is often publicly available.

  • #2
    I certainly think this is a viable option. At the very least showing the machine what is not socially acceptable will make it more resistant to picking up bad habits once it enters the "real world." As you mentioned, moving the data from one context to another may be difficult, but a clever ontologist can work with the abstractions to streamline this process.

    Working with this sort of data will yield insights. This is a tough nut to crack and we shouldn't expect to do it with one...erm, nutcrack?

    The strengths and weaknesses of this approach will help the field of roboethics advance. What's important is there is certainly a lot of data. :-)