March 14, 2008

Transparency, or Not? It Remains Unclear

by Andrew Stern · , 12:20 am

Noah’s analysis of The Sims suggests that The Sims succeeds as a game experience because it exposes the characters’ inner processes to the player. In reaction, Richard Evans, working on a related to-be-announced product, describes the debate he and his colleagues are having over how much of their NPCs’ inner workings to expose. Richard’s position is that players need “a clear mental model” of how the characters operate in order to for players to “project” themselves onto the characters — in particular, to allow players to believe the characters are deeper than they actually are, to believe in them as true characters.

This is a perfect opportunity for me to revive a discussion from about a year ago, “Transparency in the Behavior of and Interface to NPCs“. A very good discussion was just getting underway at the time, that due to time constraints I never added further comments to.

I’d like to continue that discussion, if any of you would like to. Please (re-)read that post, or my attempt here to summarize the discussion’s essential points:

I (Andrew) wrote: when interacting with a system/simulation/world, transparency is highly desirable, since transparency makes a system easy to learn, understand, and use. Simultaneously, we desire to make humanistic NPCs that, via interaction, allow players to experience and gain understanding of the nature of real people, e.g. human behavior, psychology, and culture. An essential human quality is our messiness: people are complicated, mysterious, nuanced, moody, fickle, often surprising and unpredictable under pressure. Similarly (and problematically), compelling characters are not transparent; you can’t control them, and that’s the point. That’s why they’re interesting to interact with. Real people aren’t machines that can be fiddled with once you understand their mechanism. In fact we should build our NPCs to get annoyed if you try to break them or crack them! Furthermore, exposing the inner workings of NPCs can hamper players from believing in them as flesh-and-blood characters, since their artificiality is made so obvious.

In the discussion, Nicolas H. agreed: “We can’t read minds. We can’t be in other people’s heads. … I know many Non-Gamers (especially women) who think that this is the fun in human interaction: Guessing what other people are up to, how they ‘tick’ inside.”

Breslin countered with several insightful points, with a similar view to Richard’s now. “I think it’s wrong to conceal the mechanism entirely, to try to make the mechanism too smart to be gamed, and so on. Artificiality is inevitable, there’s always going to be cracks in the façade, but the most compelling reason against gameproofing the mechanism is not the impossibility of the task. … The reason this is crucial is that otherwise you don’t get agency. Sure, you might be able to get the appearance or perception of agency, but you don’t get true agency unless the mechanism behaves according to some logic that is more-or-less intuitive or learnable in one traversal. The various interactions have to behave in roughly the same way, or the whole gameplay will seem arbitrary and capricious. … [T]here’s no need to pass the Turing test at all, and in fact it’s important to fail the test or the whole magic of the experience is lost. What you need to do more precisely is pass the Turing test for brief moments in time, to make that fleeting emotional connection, but then pull the curtain back and show for all to see that it was a trick. — And perhaps even show how it was done.” Breslin said players want to gain super-knowledge of the system, allowing them to master it upon replay; “a major part of what you’re enjoying is the superhuman mastery of and familiarity with lines which in life are opaque”.

DocMara countered Breslin, using the example of Bill Murray’s character in the movie Groundhog Day, suggesting that Murray’s character actually gets bored with his super-knowledge of the system he’s trapped within; that predictable characters are melodramatic, which gets boring. “Drama isn’t control. It’s much more interesting than that.”

Yet Gilbert Bernstein reminded us that “NPCs actually are machines, and you are controlling them. Furthermore, controlling things, that is making choices that affect the game state and express agency is a huge part of what a game is. … [I]t’s worth expecting players to view the NPCs with varying shades of grey between machine and person, possibly even laterally from there if you consider fictional characters to be something else.”

Okay. I appreciate Richard, Breslin and Gilbert’s points. Let me try to integrate them into my original take on this issue.

As you probably know, I too believe agency is of primary importance for satisfying interactive characters and stories.

I also agree that the system needs to react consistently and coherently to the player’s actions, to make it possible for the player to build a mental model of how the characters tick.

I’ll concede that it should be possible for players to gain superhuman mastery over the characters; that after replaying the game/story many times, the player can gain enough knowledge to be highly and successfully manipulative of the characters. I had said “compelling characters are not transparent; you can’t control them”; I’ll change that to say that compelling characters are not simplistic and quickly understood, but over time, they can be manipulatable.

Yet I maintain that it is very desirable to not expose the literal inner workings of the NPCs. As a designer, I want to assist the player in believing in the characters are real — though of course players know the characters are artificial, as all fictional characters in all media are.

So the question I’d like to ponder is, how to design an NPC/story/game that allows a player to construct an accurate mental model of character behavior, without “showing the data” (e.g., numeric stats and sliders) of the characters?

My belief is, it should be possible for the player to discern a character’s inner motivations and tendencies — how they “tick”, and therefore how they can be manipulated — by interacting and conversing with them in a variety of ways over time, in the same naturalistic way we learn over time how people in real life tick.

This is assuming the character is expressive enough in their dialog and action, to allow for this accurate mental model to be created in the player’s mind. Further, dramatic compression (i.e., making interesting things happen with an efficient pace) can speed up the process, without veering into melodrama, if desired.

Allowing a mental model of a character to be learned solely by NPCs’ dramatic performance is a much, much harder design and technology challenge than NPC expression via exposing numeric stats and sliders. For that reason, it’s probably too risky for a multi-million-dollar commerical game project to do more than incremental advances towards it.

Let me add: exposing an NPC’s inner workings, and thereby losing that suspension of disbelief I desire, doesn’t make the experience bad, just different.

In fact, perhaps a player could have the best of both: the option to successfully play purely by interpreting the NPCs’ dramatic performance, then getting to peek under the hood of the AI, then covering it back up again (as Breslin alluded to). It would be the equivalent of turning on or off the director’s audio commentary on a DVD movie.

In a sense, I’m agreeing with Noah when he says players can learn the system model through gameplay, but I’m not satisfied The Sims‘s reliance on expression via exposed numeric stats/sliders.

23 Responses to “Transparency, or Not? It Remains Unclear”

  1. Mark Says:

    I agree mostly, I think. This isn’t a full theory or anything, but my working model for compelling characters is that a player should be able to map a predictive model onto them, but not necessarily the same model that the character is literally using internally. Knowing the exact code-level model is boring, but having a high-level model that mostly works, expressed in human terminology, is interesting. So a player makes up stories—which are reasonably accurate—about why a character does certain things, but these stories probably shouldn’t really involve explanations like “because the slider, which is incremented by 0.2 per negative action, exceeded a threshold”. To go out on a limb a bit I might hypothesize that this is what goes on with human interaction also: I can predict people’s behavior to a greater or lesser extent, but the model I use to do so is mostly a fairly high level one that doesn’t involve me literally modeling how their serotonin levels vary over time. But this literally quite inaccurate model does support some interpretation and predictability.

    There is a bit of a difference, though, in that humans interpreting other humans are mapping a fairly simple high-level model onto a very complex low-level model. With game characters, the low-level model is usually actually kind of simple, so it’s not clear how well it supports making up high-level stories that are interesting and partially predictive.

    Sort of along those lines, Mateas’s “alien presence” idea is, imo, an interesting investigation of the edge case of a deliberately fairly tenuous/muddled relationship between how a system actually operates at a hidden low level and how it’s partially interpreted at a high level.

  2. andrew Says:

    I’m curious to hear a bit more from Richard on what seems to be a paradox in the concept of exposing the mechanistic inner workings of an NPC, in order to allow players to believe in them as deeper / real / lifelike (non-mechanistic) characters.

    I don’t disagree that the above paradox may hold; after all, people have learned to suspend their disbelief about artificial fictional characters in general. There could be a particular kind of pleasure in alternatively hiding and then revealing the inner workings.

    My guess is, you believe that by offering transparency of the characters’ inner workings, it becomes a (the only?) means for allowing players to form the mental model that is so critical for player agency and satisfaction — and that’s why you are for it.

    But if there are other implementable means by which players can learn such a mental model (as I wonder about in my post), which approach would you choose?

  3. noah Says:

    Andrew, we’re definitely on parallel tracks here. Next week I’ll be posting Expressive Processing sections on Facade — and pointing to the idea that the SimCity effect isn’t the only possible route to agency. Basically, I suggest that there’s a persistent vision of being able to have players experience agency by drawing on their existing models from other media (like theater and cinema) rather than coming to understand computational models. But, of course, that’s really, really hard. In the meantime, part of what is interesting about research in things like expressive language generation is that we can identify routes to exposing a variety of internal system states within the context of the fictional world (rather than through, say, needs meters).

    Anyway, I’ll go on about this at much more length on Monday and Tuesday.

  4. andrew Says:

    Cool! Thanks for the preview, which serves as a response to my question. I look forward to Richard’s and others’ repsonses too.

    btw, as I’ve stated in the past, and now put in the terms of this discussion: one of Facade‘s biggest failures is not doing a good enough of a job in helping players form a mental model of the characters, and player agency suffered for it.

    This discussion has already made this critical design goal all the more clear to me for future work, so, thanks!

  5. jtirrell Says:

    “[…] but I’m not satisfied The Sims’s reliance on expression via exposed numeric stats/sliders.”

    This is a good point, because it raises the question of whether sufficient complexity is distinguishable from “true” agency. But let’s keep in mind that interactions in The Sims aren’t wholly reducible to numbers (or at least, no more than any other kind of interaction) unless you limit its scope to the game code itself, and through this lens, all software is reducible to numbers. I think the key feature of this issue is that The Sims is not actually, as is commonly depicted, a person simulator; it is a situation simulator. Sims are enmeshed in a broad network that includes players, in which the distinction between player character and non-player character (as well as that between actual and virtual) is ambiguous. It is this distributed, situational agency that allows “players to experience and gain understanding of the nature of real people, e.g. human behavior, psychology, and culture.” It demonstrates that identity/agency isn’t discrete, but predicated upon a network of influences incorporating tangible and digital personas and things.

    I have an in-progress paper about this issue online, for anyone interested.

  6. Richard Evans Says:

    Distinguish between the data describing the current mental state of the agent, and the processes that update that data over time.

    The data should be completely transparent to the player, but the processes should be hidden. The player reverse-engineers the processes by directly observing the way the data changes over time.

    If the player isn’t given the data or the processes, then there are too many unknowns (to use the metaphor of simultaneous equations). In fact, it is worse than that, the player doesn’t even know how many unknowns there are!

    If the player is given the data and the processes, there would be nothing for him to reverse-engineer, and the whole system would be too manipulable.

    I think exposing the data, but hiding the processes, is how to steer between the Scylla of opacity and the Charybdis of mechanical manipulation.

  7. andrew Says:

    Richard — theoretically, can the data describing the character’s mental state be expressed, precisely enough, via naturalistic dialog and action? That’s how I’m seeing this debate. My instinct tells me, maybe. This can probably only be answered through attempts to build such a system.

    jtirrell, could you further explain what you mean by “the question of whether sufficient complexity is distinguishable from “true” agency”? Your other points about distributed agency are interesting.

  8. Richard Evans Says:

    Yes, absolutely, mental state can be expressed via dialog and action. But, given the current state of animation tech, a computer-animated character can only express at most one aspect of the complete mental state at any time. (He can express whether he is annoyed with Jim, or whether he is stressed about work, but cannot express both). Good human actors can express a number of aspects of their mental state simultaneously.

    If you’re only showing one aspect of the character’s mental state, you are not showing *everything* the player needs to build an accurate mental model.

    Some argue that it is ok only showing part of the character’s mental state. It’s ok as long as you’re showing the “most important” part of the character’s current mental state. “The player doesn’t need to know the rest”, it could be argued. But this response attempts to pre-judge for the player what is and is not important for him to know. It is pre-parsing the situation *for* him. This is restricting the player’s agency at its most fundamental level – instead of letting him perceive the situation, we are perceiving it for him, and giving him just what we think he needs to know.

  9. andrew Says:

    But, given the current state of animation tech

    I see this more of an issue of limitations in discourse management and NLG tech, i.e. the tech needed to choose what to say and when, and/or to dynamically write expressive dialog.

    I agree, it’s important to fully express the character’s mental state. For this to be done with dialog, acting and gesture, it would probably need to be performed over the course of at least a short period time, such as several lines of (dynamically generated?) dialog. Each line, facial expression and gesture would need to be packed full of information — yet even at best, it would probably be a less efficient communication of information, and certainly less precise, as a display of stats/sliders.

    Yet, aesthetically, I seek experiences that retain that naturalism — even if they are less efficient and risk miscommunication, as long as they are adequate.

  10. Dan Shiovitz Says:

    I think both you and Richard are confusing transparency of inputs and outputs with transparency of process. Lack of transparency in the former is bad, because it discourages interaction with the NPC — the player becomes frustrated trying to tell the NPC things or understand the NPC’s responses. Lack of transparency in the latter is good, because it encourages interaction — the player wants to figure out how the NPC works, which is exactly in sync with the story goal you want them to have.

    So say you want a situation where you’re talking to Alice, and the goal is to get her upset. It’s good to let the player know right off that the way to talk to Alice is >TELL ALICE ABOUT X, and it’s good to let them know that Bob is a reasonable topic to tell people in the game about. It’s bad to display “Alice thinks about Bob .. Alice considers Bob’s relationship to Carla .. Alice gets slightly more upset” when you >TELL ALICE ABOUT BOB. But it’s good to display the last bit, “Alice gets slightly more upset” — that creates an immediate question and goal for the player, why does Alice get upset? When they think they know the answer, they can do something that demonstrates their system mastery and try >TELL ALICE ABOUT CARLA and be rewarded for it.

    (That said, what I wrote doesn’t really apply to the Sims, where you’re not interacting with the characters as equals, you’re interacting with them in a God-role. In that case seeing them work their brains is the equivalent of seeing your dog fetch a ball — it’s part of the appeal of having a pet.)

  11. andrew Says:

    The discussion has focused in on the question of transparency of a character’s mental state, and how to output (express) it. I think we all agree, we want players to figure out the character’s mental processes — i.e. build their own mental model of the character’s mental processes.

    When I mentioned dialog technology as the limitation, I forgot for a moment that Sims characters, at least to date, don’t have the luxury of speaking dialog! So yes, traditional Sims characters would need to do all their expression in animation and iconic thought balloons.

  12. Ludus Novus :: Transparency on the Other Side of Agency Says:

    […] Grand Text Auto, Andrew Stern has reopened a discussion that ties into my last podcast. In that episode, I asked whether an NPC that ignored the […]

  13. andrew Says:

    This is great place to link to a recent Clive Thompson article, “Videogames are Post-Turing“, drawing heavily upon a presentation by Bart Simon of Concordia University, “Human, all too non-Human: Coop AI and the Conversation of Action” (Word doc) at last year’s DiGRA.

    Juicy quotes from the article:

    “The solo game… is about the pleasures of hanging out with machines even when you’re aware they’re merely machines.”

    “[S]lightly dumb and helpless AI can often seem more emotionally ‘real’ than stuff that’s trying to be too smart. Much like the uncanny valley effect in graphics — where cartoony characters can seem more ‘real’ than super-detailed faces — AI often seems most gripping when it hits a sweet spot considerably below omnipotence. If the AI is actively asking us for help, it triggers what sociologists call ‘interpretive charity’: We feel more warmly toward it.”

    “Perhaps most interestingly, Simon thinks that gamers actually enjoy the process of gradually understanding the logical rule sets that govern the behavior of our AI friends. ‘You have to suss out their algorithm,’ he says. We learn what makes them artificial, but we also understand them more completely — it’s the machine-age version of psychology.”

    For those at DiGRA last year (including Mary and Michael, from whom there was woefully little blogging — I tease), did you see Simon’s presentation, and/or talk to him?

  14. Game AI Roundup Week #11 2008: 7 Stories, 3 Quotes, 1 Video — Says:

    […] Transparency, or Not? It Remains Unclear […]

  15. Chris Says:

    Hm. I’m going to duck in for a moment and say: What was interesting about the Sims was that it suggested a whole different model for analyzing and interpreting personal and interpersonal relationships. It took a fairly simple metric — the status bar, which had been used for “hit points” and “magic points” before — and fleshed it out just enough to keep it still simple and understandable, but still getting it in the ball park of something recognizably human. The idea that, if you were going to posit a “friendship percentile”, and that certain types of interactions would cause it to increase or decrease, and this would in part be based on what the percentile was at, and… What was great about the Sims was that it gave you (well, me) an analytical tool that, while for sure oversimplified, nevertheless could still be usefully applied to the real world.

    Also, you got to build houses. That was fun too.

    But anyway, it seems to me that the last thing I would want in a game is a totally naturalistic/realistic expression of what is going on in an NPC’s (or PC’s) head. Programmers would have to come up with the interpretive models to implement it; it seems unfair that they should keep those models to themselves. Or, in other words, don’t try to replicate the outside world into the game world; come up with interesting interpretive models in the game world that cause us to re-see the outside world. It sounds like you’re getting a bit hung up in the artificiality of artifice, rather than embracing it for its positive, creative powers.

  16. andrew Says:

    You’re right, The Sims does a great job at representing human behavior in terms of stats and slider bars. And, as Noah’s post describes, it makes it all understandable and easily manipulatable by the player.

    the last thing I would want in a game is a totally naturalistic/realistic expression of what is going on in an NPC’s (or PC’s) head.

    Why? I can only guess, but perhaps you would dislike losing some control / god-like power over the characters.

    For my taste, I prefer peer-to-peer relationships with characters, as opposed to god-like relationships. As a player, I don’t want to have more power than the characters I’m interacting with; it’s an aesthetic preference, I suppose. I also don’t enjoy the typical “player character” layer in a game; I want to completely be the player character, that I fully define, without being told who I’m supposed to be, what I’m thinking, etc. Cutscenes are therefore an anathema.

    come up with interesting interpretive models in the game world that cause us to re-see the outside world

    Yes: drama. Drama can do that. I’m not interested in realism; I’m interested in experiences with a particular dramatic focus — life with the boring bits taken out. The particular bits that comprise the drama are carefully chosen to implement a particular model of a scenario.

    It sounds like you’re getting a bit hung up in the artificiality of artifice, rather than embracing it for its positive, creative powers

    Well, there are many ways to represent and enact interactive characters and story, and the interface to them. We’re debating the finer points of differing approaches.

  17. Borut Says:

    But compelling characters do have aspects of transparency for believability – their motivations and interpretations. If an important character in a story does something you completely don’t understand as the reader, you either expect the mystery to be revealed eventually, or you lose interest in the character and potentially the story because here’s a character you just don’t believe in, they’re doing things that make no sense to you. And this is of course specific to the reader – reading a romance where one character is torn between two loves, if you can’t understand the rationale the character chooses to pick the love interest you don’t like in the story, that’s a sharp kick in the pants for you being into the story.

    Mark, the problem with having the player have a different model of what’s going on in the NPC is that you will have to do a lot of work and create a lot of assets to communicate this other model as opposed to the actual model the NPC is using. The key is to communicate only certain aspects of the data. Unlike Richard, I don’t think there’s as much of a problem of cherry picking data (so as to not allow the player opportunity to parse the world themselves), the key is really streamlining what data is used in the game’s mechanics of interaction – what information does the player truly need? We do this in designing a game’s HUD, I think the same design skills can be applyed without limiting the player here. The state info about the agent to be high level enough that the player can reason about the processes (which what goes on in all games with respect to game state->game dynamics), but no more, no less. The characters can still have depth & believability, as long as you have a large enough abstract distance between the high level data you reveal & the lower level data used in actual decision making processes.

    I think it’s certainly possible to convey that type of state information naturalistically through dialog & performance, but it requires techniques involving consistent use of visual language & writing designed to do just that – which is what makes it fascinating endeavor, to construct animations & dialog that seem natural while at the same time conveying at least some of the bits the AI using going to use to color its action or interpretation of the world. While there are certainly techniques in other media that can serve as inspiration, we’re pretty much making it up as we go along here.

    And as Richard points out, that information is going to get conveyed over a period of time for a variety of reasons. At the very least it’ll take a line of dialog, but that in turn may get delayed because the character has to express something else, and some things won’t get expressed simply because there are more important things be doing (and then the appropriate window of time passes). All those things introduces fuzziness for the player’s reception of what data is being conveyed.

    So if you’re trying to express the data in as naturalistic a means as possible, that adds so much (desirable) noise to the signal that you’re better off if don’t obfuscate the data any other way – it’s always easy to add more complexity to the communication at that point, much much harder to make it clearer and more explicit. So I think it helps to start from that end.

    I disagree that human actors express many aspects of complex mental state at once – it certainly seems like they do, but when you actually try to pinpoint instances, it’s much more elusive. I think human actors are just much more natural at quickly transitioning though those expressions. So in that sense, I think it’s an open question as to whether it’s purely a tech limitation vs. human limitation (although there’s certainly tech limitations in addition to human ones).

  18. Richard Evans Says:

    Here’s a concrete example where we need the character to express two things at once: a terminally shy person hears the doorbell ring. His neighbor has come to visit. Does he let him in? On the one hand, the shy guy understands that it is socially expected to answer the door; on the other hand, he doesn’t like meeting people. What does he do? If we look in the debugger, we the developers can see that he is torn. But can *the player* see that he is torn? Ideally, the animation would be able to overlay reluctance over the answering-the-door animations, so the shy sim reluctantly answers the door, thereby manifesting both that he feels the pull of social norms, and that he doesn’t want to. But we don’t yet have the animation tech to express both at once. So what do we do in the mean-time? We could express it in a less realistic form (via thought-balloons or text or inner mutterings), or we could refuse to express it at all. But if it is always hidden, it isn’t part of his mental life. An inner process stands in need of outward criteria – “nothing is hidden”.

  19. Borut Says:

    Ah, but can you really approach that as expressing two things procedurally? Or is the character expressing one thing, the conflict? I mean, aside from performing physical action, emotionally he is conveying one thing – his reluctance. In part, you (the player/viewer), have to imply in some ways what he is reluctant about.

    Here’s where layering plays more of a role, directing the reluctant performance to objects in the world, giving the player some feedback as to what he’s reluctant about (and another place where some vagueness is inevitably added – did you notice he heard the doorbell? What if you didn’t hear it yourself? etc.)

    But that’s a different technical approach than say he’s going to play a performance emoting his desire to answer the door, layered with a performance emoting his shyness (each of which in turn have to layer what their source or object is). That seems much more, unnecessarily, difficult from an authoring perspective. (And that in turn means it’s easier to express scenarios like that one if the conflict state is an explicit part of the agent’s mental model.)

  20. andrew Says:

    “nothing is hidden”

    Definitely — both the character’s shyness and compulsion to answer the door need to be expressed.

    we don’t yet have the animation tech to express both at once

    I think we do… Approaching this with a director hat on — why not have the character’s locomotion towards the door interrupted by occasional hesitancies, such as stopping and looking concerned, looking away… He could even verbally express his fear with anxious sighs (not inner mumblings…)

    You’re going to need a more difficult mental state to express than that one, to prove your point… ;-)

    (Admittedly, the above action doesn’t explain why he’s conflicted, just that he is conflicted for some reason. If the why is really important, he’d have to mutter something out loud to himself, like “I don’t want to see anyone right now…”)

  21. Grand Text Auto » Link Madness, Part 2: Down to Earth Says:

    […] andrew on Transparency, or Not? It Remains Unclear […]

  22. Ernest Adams Says:

    “…we wish to create worlds as real as, but other than, the world that is. Or was. This is why we cannot plan. We know a world is an organism, not a machine. We also know that a genuinely created world must be independent of its creator; a planned world (a world that fully reveals its planning) is a dead world. It is only when our characters and events begin to disobey us that they begin to live.”

    — John Fowles, The French Lieutenant’s Woman, 1968

    Sounds like an argument against transparency and in favor of emergence… in NON-interactive storytelling!

  23. andrew Says:

    Nice quote!

    See this post from 2004 about why I think non-interactive storytellers (e.g. novelists) have a lot in common with engineers of interactive storyworlds.

Powered by WordPress