Since I have had no time to post in a month, and there have been no replies to my last post, there is a happy opportunity to segue straight into a little Q&A on the topic of games as information systems, derived from some response to the topic over at onlyagame.
So, onto the questions:
- How would one measure the bits of information in a game situation (and is it worth even trying)?
Answering the second part first - why measure the bits of information? - I think it is clear that no matter the graphical fidelity, games are still component systems designed from the top down. Therefore a reductionist approach to analysis can still work. And reductionism is a powerful tool. I think the idea is not that we want to break down gameplay to the point where we can say: the player has just interacted with information bit x, and is being presented with bit y.
I think rather that we want to be able to say that the stream of information coming in to the player has X profile at time T.
The how is quite difficult. The first step is to reduce the dimensionality of the measurement to the 2 dimensions that the player sees. But you can't lose the relational information between what is in the viewport frame and what may be around it (and the viewport frame is itself information bearing, especially in games where it represents the player's view). Thats the really hard part.
I think that I would have to hand-annotate a game with information before I could answer this in the general case - i.e. familiarise myself with my own proposal!
But to start with, everything in the game has a relationship with the player and a novelty to the player. Under these two headings, one could imagine a framework for assigning bits to in-game elements based on their relatability and novelty. I often think of FPS games like Battlefield here - there is so much detail in one of those worlds, but a player only assigns a little attention to terrain, because no matter its appearance it all behaves the same way. On the other hand, other players require a deal of attention, because despite a certain uniformity of appearance* they can all do quite different things (to kill you!). Roughly, the idea is: how can we measure the (potential) attention budget of the player?
*Which would be a personal gripe with Battlefield games - if they skinned them in Warhammer 40K designs, then you'd have a game :D
- Can game information be considered comparable? Consider the analogue information in the state of a snowboarding game versus, say, the positions of the players on the pitch in a sports game.
Very pertinent - possibly the hardest problem with this approach. Could this be gotten around by considering the possibility space of the game, and comparing on that basis? In this sense, the possibility space of a sports-like game is more dispersed (softer) than that of an analogue game. But the essential nature of the information is still the same - probability weighted relatability and novelty. Its just that that tree down the slope at time T has a much higher probability of still being down the slope but closer at time T+1, than that player down pitch has of still being in the same spot a second later (but consider the beautiful game - the goalkeeper has a pretty high probability of being in roughly the same spot! Its all degrees).
- Are information and time the only factors influencing game difficulty?
Well, novelty is very important, and is kind of assumed in the information approach. But novelty can only be judged by known play history of the player. Who can only be identified on a profile sign-in basis. Which system can only be trusted to be valid, not known. So thats a problem.
In fact, its the same old problem with player modelling again - after a certain point, without biometrics we really can't be sure that the person playing is the same one we've been modelling all along. The fields of concept drift and concept shift have methods for dealing with this, but again its all a matter of probabilities.
4 comments:
zenBen:
I admire your faith in reductionism
here, but I don't share it. :) I'm not so convinced this will yield results in a formal sense - but there might be a way to come at it tangentially.
I am reminded of a presentation at GDC two years ago. It wasn't very well attended; I remember it mostly because I was referenced in it, which was nice - but made me feel annoyed that they had rejected my talk and accepted this one! :)
What they were doing (at great expense, I might add - they had millions of pounds to spend) was breaking down the play of a game into the minutest possible details. They had broken down Halo 2 levels into nodes representing the viable action points, and had all sorts of metrics for measuring all sorts of aspects.
To be honest, I thought they were using a sledgehammer to crack and egg, that most of what they were doing wasn't going to yield results, and that they had focussed entirely on ludic agon and ignored every other aspect of play. But they seemed keen, and they were trying to keep something under their hats so its possible they had something interesting, they were just unable to show us anything interesting.
Let me say flatly that I don't think game state can be converted to an information value by reductionistic methods. However, it might be possible to measure the information load *indirectly* by some sort of biometric measure. But at this point, the whole information paradigm goes out the window, because you're measuring stress responses instead. :(
I think your proposal is fascinating, but I also think it is tilting at windmills - I just don't see how you would yield anything workable from its application.
But you might be able to explore it in broadstrokes by having very large "bins". That is, instead of measuring bits of information, you have order of magnitude categories - "Low", "Moderate", "High". These could be assigned subjectively, or quasi-objectively with some sort of rubric. Within a single game, such as Pac-man, this should be achievable.
You might then be able to get some traction, but I'm still somewhat sceptical.
I think the basic issue here is that while I agree with you that games can be viewed as information systems, we don't have the methods that can be used to complete the transition.
It's a bit like the human genome project - we can map it and make the genetic phonebook, but that doesn't actually solve any of the problems we need to address in understanding protein folding and the resultant effects - the bit we can tabulate isn't the important part. (Except, in the information systems case, we don't even have the means to make the phonebook!)
But your premise about time pressure versus information remains interesting. If you can find away to eliminate the information bugbear, you'd have a very interesting experiment at the very least.
One last thing: don't expect too many comments, as this is absurdly technical, and I'm not sure most of the blogosphere is up to this sort of scientific malarkey. ;)
Interesting stuff, though - I look forward to seeing if you can find a way forward with it!
Chris.
“What they were doing (at great expense, I might add - they had millions of pounds to spend) was breaking down the play of [Halo 2] into the minutest possible details.”
This is the PhD who runs Microsoft’s game testing lab, I assume? I think there was a gamasutra feature about his setup a while back.
“they had focussed entirely on ludic agon and ignored every other aspect of play.”
This is a worry alright. It is why, instead of jumping straight into the information metric experiment, I decided to first try to classify players based on type - using the only recent typology around, DGD. I’ve always held that getting at the root of player behavior requires looking at their motivation and predispositions.
These aren’t things you can hope to get a good idea of from reading in-game behavior in a single session.
But they are the lever for turning a dry metric of the visual state of the game into something a player model can use.
Something I haven’t maybe mentioned is that I see the information metric as an absolute value which needs to be fuzzified or interpreted before it becomes useful. Of course (as has been pointed out to me to restrain my wilder ambitions of modelling) – the player doesn’t actually see the game in terms of bits of information, nor utility functions and so on and so forth. I believe they get closer to that the more expertise they gain, but the process of information reading becomes internalized anyway. So an information systems analysis doesn’t get you a picture of what the player sees – it’s more like the frequencies of the colours in the picture they’re looking at. To ‘see’ the picture that they see, you have to have an interpretive bootstrap. I’m not sure what that could be yet.
In this sense, what use is the information metric? It seems very much like cracking an egg with a sledgehammer. We can play the damn game and get a far better idea of what it looks like to the player than such a reductionist computational method ever will.
To me, the use is the possibility of seeing further into the cognition of the player than we ever can by simply playing – because our self-insight is not codified, recorded, or very accurate (obscured as it is by the constructs of the self).
Having the ability to hold fixed some aspect of cognition means that we can look for patterns, repetition, even just hints of where and how to look next. And here is where I admit (again) my interest in games is the value they have as a test-bed for studying non-pathological thinking.
Now I’ve strayed a long way from the practical experiment, back toward the hand-wavy theorizing that inspired it. I can say that barring a post-doc research assistant position, I won’t be investigating this experiment in my current context. The end is nigh!
I’d rather do it myself than have it lifted by an avid reader (!) but in the end if it’s a good idea, let it flourish under whomever, and if not, exposure should kill it off :)
Just a quick point here, and maybe I misunderstood this part, so correct me if I am wrong. When you say:
"In fact, its the same old problem with player modelling again - after a certain point, without biometrics we really can't be sure that the person playing is the same one we've been modelling all along. The fields of concept drift and concept shift have methods for dealing with this, but again its all a matter of probabilities."
This seems to assume that a person has a static personality tied into some quantifiable value (like biometrics). The question that this point raises in my mind is, even if you could use some unique id which will apply to a person throughout their lives, a persons personality does not remain the same...so as a person is playing the game and experiencing things both within the game and outside the game this user model must become inaccurate.
So I suppose the point I am making is, even with biometrics, you can never be sure the person playing is the same person who played before...in fact, one should be assuming that the person is not the same person by the very fact that the player is accumulating a history, which in turn is molding their personality.
Of course you're right Kris. The issue is that the kind of drift that we experience in a single player, the change in their personality and experience, is a different beast to the shift that occurs when the players change.
The italics refer to terminology in the field of mining streaming data.
A concept in this field is information that has been inferred by data mining.
Concept drift is when one of these concepts changes over time, but retains attributes that relate it to previous concepts (remember here that data mining operates in a snapshot manner - so each concept is derived from a window on the data stream).
Concept shift is when the change is more fundamental. Thats about as well as I can explain it!
Search for papers by Michaela Black and Ray Hickey for more (warning: technical!).
Post a Comment