Since I began writing this column for AiGameDev.com a month ago, Alex and I have occasionally spent an amusingly large amount of time deciding on the titles for the articles. People who have seen my writing before know that I am often inclined to tongue-in-cheek or intentional double-meaning in order to illustrate a point. This column’s title is one such example… and one that I won’t let Alex edit because the question that it asks is very illustrative of a disturbing quandary that AI programmers face.
You may recognize the loaded question that is being parodied by this title: “Does this [clothing] make my [body part] look [good/bad]?” It is a comedy staple due almost entirely to the reaction that it elicits from people (mostly men) as they decide how to split the difference on the equally-damning answers. Is honesty the best policy, however? If he lies to her, isn’t he helping her lie to himself? And if she decides that she does look ridiculous, is she going to hold it against him for not saying so in the first place? We truly feel for the poor sop as he opens his mouth knowing that, no matter how he decides to answer, he’s going to wish that he had chosen the other route (or none at all). And does she really want to hear the answer anyway?
The title question can be thought of in much the similar vein. Instinctively, we want to answer “no”. It’s intuitively wrong. A blatant contradiction contained in one eight word package. After all, since the inception of game AI, we have been trying to have our agents make better decisions, not worse ones. We have been trying to eliminate stupid behaviors, not encourage them. We have been striving for more realism, not less… but wait a minute… Isn’t this where we might be lying to ourselves? And do we really want to hear the answer if it means we have been wrong all along?
A Rational Definition of Rational
John von Neumann was the father of game theory and, in a broader sense, decision theory. Most AI programmers who are familiar with his minimax model also know that one of the requisite premises of applying it is perfect knowledge of the game environment. Given that, the minimax model is spectacular in calculating the best possible move at that moment. Even when faced with uncertainty (i.e. imperfect knowledge) surrounding the game space, there exists an entire arsenal from which we can draw. Bayesian math, for example, helps us to fill in the logical gaps that may occur when we don’t know the whole story. However, this is still one shortcoming that John von Neumann, his writing partner Oskar Morgenstern, and even the legendary game theorist John Nash all acknowledged. All this fancy math to predict decisions based on maximized return is based on the premise that the decision maker is rational — which, much to the dismay of many, is not always the case.
“All this fancy math to predict decisions based on maximized return is based on the premise that the decision maker is rational — which is not always the case.”
The issue that the above-named luminaries were concerned with came to the fore in a roundabout way. Their algorithm had to generate a decision based on what the same algorithm said the other side would do. That is, “if the other player makes his best decision, it will be X, to which I will respond with Y.” However, all of it fell on its face if the other party was not rational — and therefore could not be counted on to make the “best” decision. The point is, humans were not rational — or at least not perceptive and calculating enough to determine what that maximized solution to any given problem may be. (Rational, as used here, is almost a catch-all for omniscient, perfectly calculating, unemotional, and other such non-human traits.)
One common situation would be where the competing options are not significantly different enough from each other so that it is apparent that one is the “best”. A purely mathematical approach may be able to quickly calculate the intricacies and assign a score to the options from which a rank can be determined. For humans, however, this is not always an intuitive process. Let’s face it, there are entire self-help books dedicated to teaching people how to construct a simple “pro/con list” — and that doesn’t even get into the subjectivity that may be involved with assigning values to items on the list. Purely rational — i.e. algorithmic — decision making is just not something we excel at. Often, we have to “go with our gut” — which, if not perfect, may at least lead to a reasonable decision.
But that’s not how we program our AI, is it? As I mentioned above, in modeling our decisions, we have constantly been focused on finding the best that we can accomplish. The A* algorithm finds the best path. We take great comfort in the fact that A* even guarantees that it will find that best path or our money back. We can roll up the n-depth tree search with thousands of leaf nodes into a handful of individually scored decisions, pick the highest ranked one from the pile, and feel quite comfortable that we have selected the “best” possible path. In fact, if our AI is criticized for being not human, we instinctively look for ways to make our best decisions even better! All of this is based on that the same fallacious ideal that haunted John von Neumann — “Humans are always rational; Humans always make the best available decision.”
What is Better Than the Best?
If their algorithms were to accurately model human behavior, the fathers of decision theory had to eventually accept, account for, and try to model the very fact that humans do not always act rationally. This was a daunting task that has only recently been addressed by delving into the realm of psychology. Thankfully, this field is well beyond the scope of this article. (Although I do recommend “A Beautiful Math” for light reading on the subject.) One advantage that we have over von Neumann and Nash, however, is that they were trying to accurately predict human behavior whereas we are only trying to model a sample of it.
Even with that disclaimer in force, the spectre of the imperfection of perfection still looms over us. If the optimal decision isn’t the best one, then what is? Certainly the collection of “good” decisions will tend more towards that optimal one than they will towards a purely random one. Constructing the stochastic equivalent of a blindfolded dart-throw isn’t going to be the answer. In the situation where there are multiple, similarly scored alternatives, picking amongst them could generate a unique collection of behaviors that isn’t precisely deterministic. No matter how we select things, we would want to limit the choices to ones that that are, if not the optimal ones, at least somewhat reasonable. Anything blatantly stupid only makes us just look bad. No amount of explanation of how “we were modeling a form of stress-induced psychosis” will rescue us at that point. Therefore, our “mistakes” (I’m beginning to prefer “inconsistencies” just to avoid the stigma) must be on an order of degrees rather than magnitudes.
“Your illogical approach to chess does have its advantages on occasion, Captain.”
— Mr. Spock to Captain Kirk in “Charlie X”
So, the Question is…
Again, a proper treatment of this subject would go well beyond the caffeine supply I have in the house tonight — and strays somewhat from the purpose of the column. The quandary that I want to pose here is not a technical one - that is how we could make our agents less than optimal. What I want to ask is more subjective (and therefore bound to be delightfully contriversial). Should we allow (or force) our agents to occasionally not do their best… to sometimes not select the best weapon… to once in a while not dash for the perfect cover… to typically not make impossibly swift analysis and decisions (see last week’s column)… should we allow our agents to make mistakes - and thereby look human? In fact, if our agents ever are going to look more human, don’t they almost have to make mistakes?
Can we seriously walk into a room and ask our designers, producers, publishers, reviewers, and ultimately our players, the loaded question: “Does this mistake make my AI look smarter?” Or maybe it is just better not to tell them that our agents aren’t “perfect”. That we actually designed them to be that way. It may be safer than asking the question. I know that I cringe at the thought of the answer.
One thing is for sure… if we get it wrong, they are not likely to be delicate about their answer. (And do we really want to hear the answer anyway?)
15 Comments ↓
To be absolutely honest, after you've played against most AI opponents for a matter of a few hours in a game, the flaws do not make it look better - the AI isn't at a level where it is competent enough to make mistakes, which is where the question falls down :)
But given a competent AI - of course, mistakes should happen. It could be scripted in any linear game - the AI discussing something, or kicking in the wrong door, or shooting the gas cylinder in the room...things like that add character (and depending on the opponents, could be a core part of silly gameplay).
However, there are some that strive for AI which can make mistakes. Might I bring up any Thief game? The entire sensory system is based around the AI not having perfect knowledge. Garret the master thief can hide and hold his breath inches away from the guards, and they can walk on by. They might make mistakes when investigating, or even attacking, allowing Garret to get away scott free.
Respawning enemies, such as in Unreal Tournament, while usually not programmed for mistakes, can have them added - I am pretty sure UT2004 has imperfect jumping at most difficulties, which means sometimes it misses a jump, or lands in the wrong place, just like a player would who was running around jumping all the time. You notice it because you play 10+ minute matches with them (or even them as your allies) so it's easy to see.
Commanders of armies also can make mistakes. Generally, it's more difficult to program, but I've seen the Total War games do the historical scenarios quite well - falling into the right traps (or not!) making the mistakes (or not!) it becomes less predictable, much more enjoyable, since you can usually see most of the enemy and plan accordingly, and he can do the same against you. This allows back and forth "mistakes".
Not many games allow the mistakes mentioned above however - basically; the AI is dumb because it only lives for a minute, or less. How can you make a mistake and get away with it in that amount of time? You simply can't - anything that acts like a mistake (even a blatant one) is usually perceived as simply "dumb" or "a bug". This is rather sad, but not much can be done about it, except if the mechanics of the games were changed. This applies to most of the games having not competent AI - if it were competent, it'd kill the player too often!
So, some mistakes? Excellent idea to do, but only if the NPC's are alive for more then a minute. Adds character, and randomness which humans have. For those less then a minute however, most of the time it'd be called buggy, and the AI is imperfect or incomplete for those scenarios anyway (ie; if you keep them alive for more then a minute, they'll do the same thing over...and over...and over...and ...over...). Hope that made sense :D
It definitely makes sense to me :)
I've been trying to formulate something along the line of this article for a while and thanks to it, I've managed to organize my thoughts better
Well, everything began when I witnessed a weird event in the first Call of Duty : I had attacked a bunker and thought I had killed everybody with the grenade I had thrown in. I was confidently approaching the entrance when a single soldier suddenly jumped out screaming from the shadows inside the bunker and fired at me his last rounds. I got surprised, pulled the trigger and killed him.It lasted less than a second but was a shocking experience : this gave the opponent a human face, casting a different light on a game in which, to this point, only allies seemed capable of human emotions.
This reaction was human, however, I wouldn't say it was irrational, and that's were I depart a little from the article and strongly agree with Andrew's comment. What the situation felt like was that the soldier had made a rational decision (i.e. that you could explain logically) based on his knowledge of the situation and its 'internal' (psychological) state. I'm sure this was not how CoD AI works, but the model stands : as Andrew said, the 'misconduct' appeared because the knowledge was partial but even more so, it was analyzed against a filter linked to a certain state of the agent : in another situation, the same knowledge would have been analyzed differently.
In the end, the agent still took the best decision but not objectively, subjectively. So, for mistake to have impact on the player experience, the game need to convey the reasons that lead the AI to the decision : in my case, I noticed afterwards that the guy was the sole survivor a massacre, that he had a few bullets left so his decision made sense.
From this, I think Andrew's point about short-living AI can be generalized. The psychological states that 'filters' the knowledge
should only be modified in ways that fits the character : cannon fodder can go from 'reasonable' to 'freaked-out' because they just saw one of their comrades die, but an elite-super-trooper or a mindless peon of a multitude shouldn't.
In any case the 'failability' of AI driven characters seems an underrated mechanisms to me : it gives depth to nameless characters, brings more credibility to the world, enables to the player to study and exploit these shortcomings - feel intelligent afterwards - and be even more amazed when really smart reactions are displayed.
I remember making an Othello game as a friendly challenge a while back. It was supposed to play fast, and good. Now as we all know Othello is a well understood game, and you can search the entire tree if necesarry, so the other programmers did exactly that, they tried to figure out how to model the perfect game in a recursive tree. And they made solid near-perfect robots.
My solution was to play Othello(I was a newbie and had never actually played Othello before) for a while and notice some basic patterns, such as relative values on the board, then what I did was a simple weight based GA (for an 1/8 of the board, othello is a mirror game) that used a few solid rules any expert would give to a beginner and had it play a few hundred (note that it wasn’t thousands) of games against a random player(as in a player that was not rational at all, it just placed it’s pieces randomly).
Then I made up 12 personalities based on different weights to the GA parameters, and voila I had 12 quite good Othello players (not perfect, but they surprised the Judge with their unique styles and non-robotic gameplay.
More interesting though was that the Judge whois an Othello expert, beat all other AI’s perfect games by twisting playing to their perfection and predicatability, but failed one game against my best (aggressive) AI’s because it wouldn’t fall for his traps :)
However on the other hand this has been the only time I had success with GAs to create a smart and interesting AI, one point of importance might have been to have given it a solid base for good gameplay before sending it out on it’s own, and the other point might have been the random challenger that assured the GA could react ok to any stupid situation, not just the rational ones.
Anyways I make RPGs and Tactical games, can’t give any solid advice on applying these ideas to an FPS
While talking with friends, I find that usually gamers don't value in the same way a mistake from a person and a mistake from a computer. For example, in FPS games, it's desirable for the AI players to miss shots (in a believable way) because if the AI always scores a head-shot it becomes frustrating. But if you find a human player that always scores a head-shot, although you find it frustrating, it's more a feeling of "oh my god, that guy is leet" rather than "the dammed computer is cheating on me".
But problems in movement (wrong navigation, missed jumps,...) mostly always look silly on AI players, while we are quite forgiving when another human does them.
Or that I have found when talking with people that aren't programmers (just gamers). Regards!
I think there is something valuable in this question.
I do think that a lot of this will come down to the psychology of players and the internal psychology of the AI's to some extent.
What we dont want to see are situationally "dumb" mistakes, like pathing into a wall, or jumping of a ledge thats too high and getting killed, or simply not reacting to a threat thats stood next to you.
Once you have these "basics" covered, then I think we're into the realm of creating believable and interesting behavior, which is where the real meat of AI lies and is where its fun for me.
Clearly some sort of simplistic internal pyschology would be a useful starting point as its that which drives most human decisions, so modelling it in some way seems natural.
So maybe one thing we have to do as AI programmers, is study psychology to learn the two sides:
Of course we're getting into the realms of human thought, which is probably quite complex, but even a simplistic model must be able to offer us something useful.
I suppose its a lot like the old sims "needs/desires" thing. A set of variables which can be satisfied by actions. They act as an ultra simplistic set of samples that can control a motivational model (sample A is too low, go satisfy that need kind of thing).
I wonder what kind of motivations would be useful across a variety of game types.
Humans for sure make mistakes.
But at the begining, we should recognize the difference between the flaw and mistake.
One being a malfunction due to poor implementation and other being a change in operation due to unsufficient knowledge or improper judgement.
How would one introduce diversity of, lets say pathfinding.
Is the optimal path really the one a person would seek ?
Will a person always take a same path (for years to come)?
Not everything is always rational in our decision making procedure, on the contrary, one takes a different path to job just for the sake of taking a different path, or because it is not the shortest, or because it is not so dull.
And after all, some people make lethal “dumb” mistakes such as jumping of the ledge into oblivion.
If they leave a note, other might percieve their internal mood.
But if they don’t, others will never really have the knowledge why they did it.
So my answer to your question: Yes.
Make mistakes, if they are not flaws.
Make diversity, that will bring you to your destination. (not some stupid attempts at climbing the walls)
Otherwise it becomes dull after few attempts of behaving exactly the same.
Thanks, as always, for making the weekly Developer Discussion column into a discussion, folks.
As I alluded to the in the article, the mistakes that our agents make would have to be very subtle. In fact, many of them might not be recognizable except as levels of depth. If 3 competing options for a decision score within a few percentage points of each other, is it really a "mistake" to select the 2nd or 3rd rather than the first? However, our algorithms are generally constructed to take the highest ranked one regardless of its margin of victory. To the player, however, the effect is that it reduces the monotonous predictability - "When in situation A with parameters B and C, the AI will always do X." That sort of determinism isn't something you get out of 100 humans (or 1 human 100 consecutive times).
If we were to use obvious mistakes, however, there ARE ways that we can "sell them" to our players to show that they are NOT bugs or errors. For example, if an AI agent decides to run for cover, (spinning off of last week's example), it may select a poor cover, get most of the way there and - with frantic animations to support it - say "Damn! That won't work!" and look around for another place to go. How about selecting a less-than-optimal spell to cast? Or weapon to use? The AI can make a comment about how it was the wrong choice. If it lives long enough (good point Andrew) it can perhaps back up to discard the weapon and select another.
There are plenty of ways to make our mistakes look intentional (doesn't that just sound odd?). The question is only if we are willing to use those mistakes to create the "humanness" that our AI needs.
Those kind of rare occurrences would work excellently for non-combat AI. For instance, if The Sims had characters making mistakes (I've no idea if they do) then you can see over time perhaps learning from the mistakes (ie; cross that off the list of things you can do wrong now).
And for reinforced learning, Black and Whites non-combat AI which did mistakes, lots of them, until being told not to do something (and I don't mean eating villagers, which depends on your morality, it can try and eat rocks until you tell it not to!). Works excellently for these models (and the Thief example) due to not having time constraints which combat commonly does.
Notably, the example of cover - those NPC's which need cover to survive? The PC would likely kill them before they saw the NPC made a mistake about going to the wrong piece of cover, since it'd be an obvious target - the only way the poor guy would survive is with more health (and thus longer life) or if the player was, say, reloading but looking out from their cover. It's hard to justify the work if it's unseen of course, so in combat, it is possibly the most difficult time to make them more human (unless their your invincible allies or something).
As noted; board game AI (or for most computer games, tactical turn based or RTS AI) can have it work with personalities - the mistakes might be intentional (or, in the AI's eyes, none existent).
As FEAR showed us (it got rave reviews for it’s AI), being able to convey to the player the AI’s “thoughts” can be more important that the AI actually having those thoughts.
As Jeff Orkin said, the FEAR AI didn’t have logic for carrying out a flanking manoeuvre, but it did have the ability to yell out “Flank him!” (or similar), which when combined with a buggy search-for-cover algorithm actually convinced a lot of players (and reviewers) that the AI was really good at flanking!
So in this case, flawed logic for taking cover (such as having to run behind the enemy to get to the cover location) combined with the smoke-screen of good communication with the player leads to many non-existent behaviours being observed.
As for the problem of always choosing the best option - I usually avoid this by adding a small random number to all of the AI’s decision weights. This way a weight of 89/100 may be chosen over a weight of 92/100. This even works well with A* to get varied, but still ‘correct’ path-finding.
Perhaps instead of using a fixed random number range (e.g. 0/100 to 5/100), the range could be based on a function of time that has occasional spikes - this way usually you’ll get a small amount of variance, but on occasion the AI will do something totally outrageous, like ignoring cover altogether and going in for the melee kill. If this occasional outrageous decision was combined with the right ’smoke screen’ (e.g. voice overs), then the players will likely see it as a very human reaction to a particular situation, instead of the pure random chance that it really is!
I agree with most of these opinions: Mistakes are only useful in an AI *if* the player notices the mistake and *if* the player understands that the AI’s decision process made a mistake, rather than the programmer generating erroneous code (running into walls being a prime example).
However, I have to politely disagree with Andrew that a short-lifespan AI doesn’t warrant “mistakes” being put in.
Let’s pick a hypothetical example: A soldier AI wounds a player. The player retreats into a different room, heals back to full health, and the solider charges in, since it’s list of instructions say that a wounded player is an easy kill.
If I was the player, under ordinary circumstances I’d say the soldier has a simple programming: I interpret the AI is just attacking me and doggedly following me in an attempt to kill me. However, if I am presented with other soldiers in a range of circumstances, I may realize: Soldiers pull out all the stops when they realize I’m low on health in an attempt to bring me down ASAP. These attempts will fail if I can fool them: pretend to have low health, or to hide clues that I’m vulnerable.
Effectively, using a single AI type throughout a game/simulation is, in fact, like giving it a long life, even if each instance of an encounter with a soldier (or other AI) only lasts a minute at most.
If different AI guys repeats those mistakes "enough times" they will end looking silly even if they try to justify them (specially coming from different entities). And I fear that for a player "enough times" is a really low number (but I don't have anything solid to back up this claim :( ).
Also, players tend to make mistakes in different ways, it would help if each AI entity made mistakes in each own ways (giving them personalities as Andrew says). I think it gives more feeling of "real intelligence" having 2 different entities doing 2 different mistakes about the same thing rather than doing the same one.
I have just remembered a sentence from a friend game developer from Spain: "in case your AI is not very clever, make the enemies zombies, player won't matter much if they do dumb things, zombies are supposed to be dumb :p". It's fun how changing the visual appearance of the enemy can make the player think differently about how the enemy should react and work (I bet that if F.E.A.R. enemies were zombies they would have seemed too clever).
Regards!
On a bit of a tangent, but arent zombies getting a lot faster/smarter these days?
Erm, anyway, back on topic.
I think there is a good deal of research that needs to be done on player perceptions of AI agents before we can really say what works and what doesnt.
Only, you would need to have:
An AI expert
A psychology expert
To do that kind of research. Not sure I know anywhere that has the two together.
This is the kind of work that would quite likely help improve the AI development process though. Maybe I'll propose it as an MSc project.
Just want to give my contribution to the discussion from a bike racer point of view. In this case mistakes are often part of the design. They can easily be one element that characterize a given pilot.
The challenge here is more about giving enough control to designers and making them the more realistic than possible.
Think there are mainly two different approaches: the first is building an AI enough competent to drive but who may fails. The second is building a perfect AI (maybe cheating a bit) and forcing errors on it. In the first case we'll probably have more realistic mistakes but few control, while in the second case exactly the opposite ...
Wow, a lot of really good points already made. Still, what really stood for me was this from Zoombapup:
”
* How the player percieves the play of the AI
”
I confirme, mistakes make an artifical character more human, but if they don’t come across as mistakes, they’re pointless and confusing for the playing experience. Adding a little help from sound/animation might already do the trick: “I know where you are, I’ll get you little piggy” > wanders off into a dark corner….
Maybe something along the lines of “human” to add: A character that actually doesn’t always do something meaningful, but rather only _seems_ busy might seem more life-like than a restless NPC. Just remembered Yorda in Ico, most of the time standing around completely clueless. This made her much more human, as in “shy and scared” than that annoying girl in HalfLife 2 that seemd to always know where to go :)
The way I see it is that those mistakes are nothing but another feature to add to the character. In the same way as we try to add variety to character by having different animations for the same action, or a whole bunch of different attacks even though there is one that is the more effective and if we want the AI to be optimal it will use always this one.
In essence I think that when you are programming an AI a mistake is not a mistake per se, it's just something you consciously put into the system as another behavior to choose from.
Any type of variety that we can add to a character will increase its believability. I think we are far away from getting perfect characters in games. I think we will continue adding "mistakes" in the sense of non optimal attacks because it is required for gameplay, but the extra time we have is probably better used in solving all the "programming mistakes" to get flawless characters.
Leave a Comment
You can also reply to this thread in the forums.