** While writing this the type of game I have in mind is primarily FPS games, but I think it also applies to other genres as well. **
People don't like what they don't understand. Being bested by something that you don't have the slightest idea what they're thinking is frustrating. What if you thought a human, something you could relate to, was behind the controls? Would you feel as frustrated by being beat? Or a better question - would you feel cheated?
Let's imagine for a moment that the perfect AI was created. Not only did it have aim similar in reaction time to a human, but it also could adapt and learn from what the player did. Would this really make the game better? Any time the AI succeeded in besting the player that player would feel a little bit cheated even if the AI appeared entirely realistic. This is because the player is knowledgeable that they are up against AI. Slant that perception so that they now know no better, they think they're up against another human. If the game played out the same exact way the player would feel very different about the outcome. They would accept the fact that they got bested and devise a way to adapt and win. In the case of going against an AI and losing they will not be as motivated to adapt due to not understanding their opponent's thought process. How can they possibly devise a plan to adapt when they have no knowledge of how the AI is arriving at its decisions?
When facing AI instead of a human opponent, not only are their actions on trial, but the method of reaching those actions is on trial. With AI you can't be sure of how the AI is reaching its decisions so anytime you lose it will be looked at as unfair. With a human you can relate a decision by your opponent to your own experience and thought process as a human and understand how a fellow human reached a certain decision, thereby eliminating the perception of unfairness.
Think of it this way: if you were to set up a bunch of dummy servers for some competitive game that were running bots (possibly even with our level of AI technology today) while trying to maintain that perception of a human vs human match, with things like fake trash talk and links to real gamer profiles - do you think, given no obvious glitches occurring in the AI, that the player would be able to tell the difference? Probably not, and it is due to the perception that what they are fighting is human. A player could probably play all night against these computers and not know the difference, but tell them that they're playing against AI and the complaints of unfairness will rise while the average player playtime will drop drastically.
Perception is the most important thing to creating convincing AI. Since AI is usually for single player games it becomes a bit problematic in creating this perception where the game isn't supposed to have other human players in the first place. Below are a couple ideas to creating this perception, or getting around the pitfall, for single player games:
- Make a gray area where the player can't be certain humans are in their game or not. A system similar to Demon's Souls where the game is primarily single player, but also has human invaders that can come into your game. Except make this invasion more covert than Demon's Souls did.
- Make the AI smart, but not good. The only time a player will question the AI is if they're losing to it, so design the game to have a minimal amount of player failures from direct AI involvement. At the same time exhibit smart looking AI, but instead of offensive smarts create defensive smarts.
Without changing the player's perception I don't think any amount of AI technology will change a player's perspective on the experience they get from a single player game. As long as the player is knowledgeable of the existence of AI, that will be enough to dissuade them from any illusionary efforts by the AI designers.
|
I've usually noticed the opposite. Against AI, humans usually figure out exactly how their opponent thinks and from there its dominos. I remember Mechwarrior 3. That game is just plain unplayable to me now; there's no challenge because I know I can just walk around the enemy and shoot inwards and it can't do a thing about it mentally.
The "AI" would have a preset goal, and would just learn as it goes along to achieve a goal. If the goal is to survive and kill an opponent, then at first, going up against an AI would be simple, but as a player progresses it will get more unpredictable. I have seen this done before in other games, but again I believe it was still programed to to take evasive measures rather then learning it on its own.
I believe there are some non game examples out goal oriented Ai learning, but im not a programmer so I really don't know if implementing a system like that would be too much on hardware resources.
Yes, but they exploit the AI to be able to beat them. If they weren't aware it was AI in the first place their thought process about how to beat it would be different, leading to better AI just by changing player perception. As for Mechwarrior that would fall under the, "given no obvious glitches" category I mentioned.
@Dan: To me, you're talking about Artificial Evolution in A.I. and that's something I've been very interested in since I read a book by Wired Editor-In-Chief Kevin Kelly called, "Out Of Control"; where he starts with observing that our current pace in the scope of system development will soon overtake our ability to manage such development, even with large teams. His idea is then to follow what some researchers have done in the past and develop simple structures that mirror natural evolution to "grow" future applications without having direct control over its development. This will sound absolutely nuts to engineer readers, I realize. The implications for games and game A.I. in particular, however, are nothing short of amazing. I've conducted tests in A.E.A.I. myself and the results have been encouraging to produce a progressive challenge to the player. In fact, I recently re-created those tests in Flash, as an A.E.A.I. toy, and plan to publish it on my website as soon as my plate clears up.
I feel the subject matter of this article alludes to the bigger challenge of finding difficulty balance with in a game. Finding that sweet spot where players don't win all the time, but still feel like winners. Or better yet a player continues to loose, but feels they can actually win with practice.
AI fails when it is rigid and predictable. Once a counter-tactic is discovered, the AI will lose every time. Therefore, even a clearly stupid decision can be a valid and interesting tactic, if it's made infrequently enough that the player doesn't expect it. In the absence of learning and evolving AI, adding to your "best practice" code an arbitrary/random tactic (that kicks in only occasionally) is one of the more cheap/cheerful options.
Another option is to investigate the field of Artificial Stupidity - the potential for machines to make tactical mistakes in such a way that the game becomes more interesting.
@ Glenn Storm
Thanks, I didn't know about the turing test, but you are right what I propose is very similar. Although the turing test appears to be testing actual intelligence or execution while I focus more on the experience the player (interrogator) gains when they don't know that they're supposed to be finding an AI. I think if they (player/interrogator) don't know that they're supposed to be finding an AI, then a lot of suspicions from weird or overly accurate behavior are disregarded more easily.
It's important to recognize that this interaction between player and (potential) bot takes place in a game context. As a game, players expect to have a chance of winning. That means bots don't have to be capable of reasoning or learning or any other aspect of intelligence; they only need to present enough of the appearance of those behaviors to provide a satisfactory level of challenge.
I think this is the point Blake is trying to make: it's the perception that matters. As long as the presentation is plausible through gameplay, it doesn't matter what is or isn't going on behind the curtain.
That actually makes life easier for developers. Although it could be very interesting (I'm inclined that direction myself) to browse doctoral theses and _Communications of the ACM_ for the latest AI research that might be used to create self-improving automata in a game, it's not necessary. The actual implementation of a bot's behavior can be the cheapest, quickest, dirtiest hack imaginable and still be a win as long as most players of the game perceive the behavior of bots as sufficiently intelligent to be an appropriate challenge.
So perhaps the more useful field of research for folks developing bots is not pure AI, but psychology: what makes people appear interesting enough to want to interact with them?
Note: There are some other interesting subjects related to this one. I'd enjoy seeing some blog posts on things like:
1. Do most players of these kinds of games truly want the level of challenge that real human opponents can provide? Or do most gamers actually prefer perfect predictability in their opponents so that they can always feel confident of being able to win in selected engagements?
2. Do bots really need to learn, or is it sufficient to spawn multiple bots with a wide range of preset ability levels? If learning is desirable, what are some effective ways to implement that? I like the notion of using genetic methods to "breed" better bots (http://flatfingers-theory.blogspot.com/2009/04/breeding-better-npc-opponents.htm
l), but what about other approaches?
"They would accept the fact that they got bested and devise a way to adapt and win. In the case of going against an AI and losing they will not be as motivated to adapt due to not understanding their opponent's thought process."
@Blake: Have you actually studied or experienced this effect anywhere? It sounds like you are suggesting lying to the player (by putting on a show for them to convince them they are playing against humans) to enhance their single-player experience. I'm not sure I understand your argument -- a player's expectations of a game are initially very different when playing single player as opposed to multiplayer, because they already KNOW they are going up against the game.
I don't see how thinking that I was beaten by a human instead of a bot enhances my single player experience at all, in fact it would just be confusing to me because I expect that the AI enemies were designed in a particular way to advance the narrative, i.e. the end boss is always in the same place because that's where the end of the level is... if the game was trying to pretend that the end boss was a human, then it could (and should) appear anywhere within the level at any time, but that wouldn't make any sense.
So let's say your goal is to create a server where human players can go expecting a multiplayer experience, but instead it is always populated by bots (presumably to ensure that player's are always given an opponent whenever they choose to play) -- where do you draw the line on your human pretense? Should bots be campers, or generally try to annoy, pester and shout profanities at the human players? These are reasons many players prefer single player experiences in the first place.
Well the closest true single player game that I've seen come close to this is Demon's Souls. Other players can invade your game and can be anywhere in the level. That game warns that your game has been invaded though. Also on one level of the game if a person invades the game they have a random chance of being able to control the boss. The boss is limited to the boss area and has to wait for the player to arrive, so your boss concern could be handled like that.
Also the point isn't to bring the multiplayer experience to single player. It's to improve AI without actually improving the AI algorithms. By hiding the AI you take a level of player criticism or skepticism out of the equation. If a player is up against the best AI there will always be the excuse of, "well it's a computer", or "it cheats".
In a way predictable AI, or simple AI as Tyler called it, is doing what I propose in a different way. For a top down shooter if the player can always predict enemy patterns it puts the control back in the player's hands entirely, eliminating skepticism as to how the AI beat them. For more realistic AI this is a little tougher because one stray bullet, grenade, or rocket that kills the player will invite this skepticism. Also many FPS or similar real world based games rely on the idea that the player is up against unpredictable humans.
So this is really a theory as to how to accomplish the feel of that simple/predictable AI without making it simple or predictable. I don't think any number of algorithms to improve AI will ever erase the skepticism that a player will have toward advanced AI. However, I think skewing perception of what they're up against can go a long way to erasing that skepticism.
As far as 'interesting', one review of GTA 4 brought up an interesting point - in MP modes, the reviewer often had more fun playing with bots rather than players. The reason was simply because the bots did more interesting things like engage in firefights while on hanging girders or what not. Players, being pragmatic, did effective things that weren't as interesting as far as impact.
On same note though I also remember an AAR of a game of Galactic Civilization 2. In this particular game, one AI opponent was basically manipulating the player and the other AIs - he wanted to prevent galactic peace (because it would give one of the other AIs the win) but also did not want to destroy all the remaining opponets, of which the player was one (because the player being destroyed would mean the end of all hostilities and thus... world peace). The full realization of this though was no apparent however until after the fact, likely because most of this process was invisible to the player other than the, at the time, unusual choices the AI made (declaring war on the player, then sueing for peace before the player could be completely wiped out).
As far as the article, perhaps the mental difference is that AI is a challenge where as players are a competition. Maybe adaptive AI should be set up in modes that aren't challenges as SP modes are but in the context of competitions as MP are.