Just to be clear, I really like Company of Heroes. I have always been an RTS fan since the original Warcraft and progressing through Starcraft, Age of Empires/Kings and especially Empire Earth. That being said, CoH is very well done.
![]() |
![]() ![]() |
|
IA Information
Communication Dave Mark's Book |
Post-Play'em - Observations on Game AI Posts Tagged ‘cheating’Odd Glitch in Company of HeroesThursday, February 11th, 2010Just to be clear, I really like Company of Heroes. I have always been an RTS fan since the original Warcraft and progressing through Starcraft, Age of Empires/Kings and especially Empire Earth. That being said, CoH is very well done. I only recently got around to grabbing the full version of the original CoH and the expansion, Opposing Fronts. (Thank you Valve weekend sales!) My prior posts here were only about the downloadable demo. Obviously, there are some significant differences in the full release version. I like them all.
Rather than continuing through the script-fest that campaign maps usually are, I have been playing a lot of skirmishes against the computer. I find a lot of the tactics by the computer to be pretty familiar for RTS games.
For example, as I mentioned in a prior post, it is obvious that there is a lot of influence mapping going on. If I find myself a little behind on producing units, I will often be drawn into a series of dashing from defense point to defense point as the computer picks on whatever is currently the weak spot. My only question in this situation is how much is the AI cheating with knowledge about my force disposition. I know that it is not that easy to discern what the computer has placed and entrenched… it seems like he knows where my weak spots are even if they are deep within my territory.
Anyway, as I was playing a skirmish map, I ran across something very odd. Normally, I had been playing against the “normal” level of difficulty. I had been winning fairly regularly against that level, however, and decided to ratchet it up one level. I was playing the Lyon map against the computer on “hard.”
Because this was a control point map, I was quite used to meeting the “normal” AI at the river as we battled for those 3 points. I figured that the “hard” AI would be just as quick to get to the river. However, I captured the first point… then the second… without seeing the AI at all. I began to suspect that either something was dreadfully wrong or that “hard” used a different tactic than “normal” – such as “build up and crush him” rather than “dash for the control points”. Not too long after that, it became obvious that something was indeed seriously amiss.
Thanks to the wonders of replay files, I was able to go back and witness what the computer did… or did not do. A video is included below.
I sent the video privately to my friend and colleague, Chris Jurney, who did much of the excellent AI work on CoH. He thought it was amusing as well. His speculation was that the Lua script exploded somehow. It obviously didn’t cease working entirely because we see some orders trickling through at times. He invited me to send him the replay file so that he could forward it on to the Relic programmer who is responsible for it now.
It doesn’t seem to be a problem with the map itself as I did go on to play a new match on that map on “hard” and it proceeded to hand me my buttocks as I expected. On the other hand, it may be a problem with a combination of that map, that difficulty level, and a randomly selected tactical approach that knocked a variable out of bounds or something. Regardless of the cause, this is a great example of how important testing is… and yet things still slip through.
Anyway, I will try to do more video captures of some games later – especially the view that shows how the computer is moving throughout the map during the game. That has been very interesting to analyze because it shows how the influence mapping and even the pathfinding is working as part of the decision making.
Doom 3: Pretty Wind-up ToysThursday, December 13th, 2007I dug out my Doom 3 again specifically for this blog. I hadn’t ever finished it when I started playing it about a year ago. It sure was purty, but it just didn’t engage me the way that I was hoping it would. Anyway, my thought as I returned to my saved progress was “pay attention to the AI so I can review it.” As I played, I became a little unsettled – not by my progress in the game, but by my progress in my quest for AI. I finally had to come to a realization that seemed almost blasphemous or sacrilegious.
Doom 3 doesn’t have AI.
Sure, I know that statement is a little over the top. Let’s face it, the AI in Doom 3is better than that of games from when I was a kid in the 80s. But that is also the heart of the problem. Doom 3′s AI isn’t much better than what was in the original Doom in 1993. Considering the related concept of Moore’s Law, you would think thought that AI would have increased at a geometric rate similar to that of anything else in computing. It is with that expectation that I claim that Doom 3 “doesn’t have AI”. Looking backwards down a Moore’s Law curve of AI, one could make the statement that the AIs in the past approach zero… therefore being zero for all intents and purposes – especially in a relative sense to where they should be today. So… my statement is now qualified somewhat. (Why do I feel like I’m going to end up having an uncomfortable beer with someone from id at the GDC in February?) The comparison to the original Doom (or Quake, et al) is not far from the truth, however. The state machine has to look something along the lines of:
That’s it. In the newer versions, you can splice in something towards the end that involves a side-step. The bad guy marines especially will do this. To me, the player, this makes me do the very involved strategic and tactical process of… uh… re-aiming my gun. Wow. Admittedly, there are some baddies that give me more fits than others because they are not approaching me in a straight line. Lost Souls tend to zip around like gnats once they get close to you which is mildly annoying and Cacodemons will float around as you shoot them – but that means they have as much AI as a helium-filled piñata. There are other differences as well. The half baby/half fly Cherubs will back off a bit, prepare a moment, and then leap. Of course, so do the little spider things. This is a major let down. I remember seeing some in-game clips at a game conference during an interview with Carmack. (I believe it was at E3.) The scene where the pink baddie tried to beat down the door, gave up and instead came crashing through the window was terrifying! “Wow! It gave up and used the window!” I told myself. When I played the game… and that exact same sequence happened, I realized that the whole thing was scripted. Oh. Bummer. Again, it wasn’t AI. Unlike F.E.A.R, where the agents do use their environment and re-plan when faced with obstacles such as the door above, the Doom solution was a movie crafted for my benefit. I have been flanked by the computer, however. This I admit. Often! But that has nothing to do with the AI and everything to do with the placement of triggers. In true horror movie fashion, the computer does come up behind you often – but only because a designer placed a trigger on the floor that says “when player gets here, spawn evil dude behind him”. That isn’t AI. It isn’t even fake AI. In fact, it’s getting rather tiresome to know that every step I take is likely to Really, the movement and attack logic for the enemy is not much more advanced than the little table-top wind-up toys. They will chatter away in my general direction, but there isn’t much purpose or reactivity to their actions. That gets very disappointing. There does seem to be a bit of cheating going on, as well. The first time I fought a Mancubus, I was doing my best to flank him around walls. Despite having moved at least 60° off to the side, when I poked my nose around the wall to take another shot, I found he was already facing right at me. The explanation would have to be that there is omniscience on my position. Here, I had decided to do a simple, but what I thought clever, evasive maneuver only to have the game say “tough… I don’t care”. It’s deflating to have that happen. There also may be some cheating going on for my benefit as well. This doesn’t fall exactly into AI, but it is relevant nonetheless. I swear that I take less damage per attack when I am almost dead. I haven’t done the math yet, but it seems that I can be in the teens or single-digit health for far longer than I spend in any other 10-number range. A hit that takes me down from 100 to 80, for example, may also only take me from 15 to 10. The result is that I spend a ton of time between 0 and about 40. Of course, if this is happening, this is a great device for creating tension in the game – and is a variant on rubber-banding. It’s rather artificial and arbitrary, though.
I have always respected John Carmack for his vision – and the “visions” that his games have provided for me. I just can’t help wondering what would happen if they were to put a bit more focus on AI. I notice on the Wikipedia site that Johnathan Wright is listed as an “AI Programmer” but that in the credits for Doom 3, he is just listed as a “programmer”. I would love to sit with him at a GDC party and have that beer. “What’s up, dog? Is there anything else you could be doing?” I can’t blame him, necessarily. Given the depth of the graphics rendering, he may very well just be out of clock cycles. *shrug* I hope that he has a better chance in the future. I would very much like to see an id offering that does “speak to me.”
Until then… there’s more than enough to challenge me elsewhere.
Call of Duty 2: "I saw that grenade coming!"Friday, November 30th, 2007
In this case, the landing area of the grenade is roughly calculated. If the projected landing spot is near an agent, they react appropriately. They yell out warnings and move away if necessary. However, the programmers would have faced a quandary as to when the grenade itself should send a message. If you wait until it lands or bounces off of something, it would discount the very possible scenario that they could have seen it flying through the air in the first place. If it is triggered when it is thrown, it creates the issue I observed above – that they know it is there despite the fact that they should not be able to see it at all.
Another potential workaround would be to create a plane that is the intersection point of the path of the grenade and the visible area of the agent. At that point, the “see grenade” event can be triggered once the grenade passes that point. That seems to be not much of an improvement since there would be as many planes as there are agents and the grenade would now have to poll the planes many times per second. The overhead would be similar to the first example.
(More observations of “Call of Duty 2″ for Xbox 360) Call of Duty 2: Omniscience and InvulnerabilityWednesday, November 28th, 2007
Enemy Invulnerability: I was on the level where your convoy through the town in Tunisia gets ambushed. I had to start it over and over because I was getting pelted by the guys on top of the walls. I finally got down where I should hide and who I should tag first. One thing kept bugging me. There was a string of guys appearing on one wall in order. Most of them were gunners shooting at me – but there was one that had a Panzerschreck. I would try to shoot him over and over and he wouldn’t die until he had fired at one of the tanks/trucks in my convoy and blown it up. In the mean time, another one would pop up and take me down. I realized that it was pointless to try to kill this dude at all. He wasn’t going to kill me and that tank/truck was going to blow up anyway. It was part of the level. What annoys me, however, is that it took me a number of times before I realized that, until that guy got his shot off, he was invulnerable. As soon as he was done with his job (that I was trying to prevent) he was now able to be shot. I wasted valuable time and got very frustrated by the fact that the level designers had decided that this was such a required series of actions on the level that they would break the rules.
To me, it seems like this is a case of the “anti-sandbox” concept. Sure, a somewhat linear game like CoD isn’t meant to be a sandbox – and there are certain things that have to happen to advance the plot. That’s fine, but I always feel cheated when I can’t change the series of events even if I do the right thing to disrupt it. Enemy Omniscience:
A possible solution to this is to measure how much of me is visible and then combine that with a coefficient based on how far “off center” I am from their current direction of vision. Perhaps another factor based on movement. I know that is a bit expensive to calculate for each AI (since their fields of view would all be different). Another potential solution is to cast multiple rays to different parts of my body – perhaps shoulders, head and a couple of lower torso spots. If more than ONE is visible, then I can be seen. Not knowing exactly what mechanism they are using, it’s difficult to know how to improve upon it. This is even more alarming when it is obvious that there is a bias towards firing at ME. I may have 10 squad members all hidden behind objects and taking pot shots at the enemy – or even running around in the open, but damn it if the AI doesn’t want to fire at me instead. It doesn’t happen all the time, but it is obviously biased more towards me than my squad-mates. This is an obvious design decision to make it more exciting. I understand that. However, when combined with the omniscience above, it’s kinda creepy. [note: I've realized that waiting to write one complete writeup on a game is sometimes prohibitive - so I will write as I think of things... and tag them by game so you can find all relevant stuff] Mario Kart: Double DashFriday, November 9th, 2007
Anyway, many years of playing the game have taken another toll on me – I have not been able to stop myself from making mental notes about the AI and the game design decisions that went into making it tantalizingly addictive. I even mentioned as much to Steve Rabin, Senior Software Engineer at Nintendo of America, while we were having a beer at one of the ubiquitous evening fests during GDC in 2003. If you have ever met Steve, you know that he would make for a fantastic poker player in his ability to not show a reaction. Therefore, I have little idea if the observations and conjecture I made about the DD AI at the time were even remotely close. I would like to think, however, that there was a twinkle and a smirk. Of course, I have no reason to believe that Steve was even involved in the development – perhaps he was just amused at my enthusiasm. Anyway, here are some of the observations that I made over the past few years. (Has it been that long?) Forgive me if I get either too simplistic or too esoteric… The State of Racing… One of the obvious characteristics of the AI is its use of FSMs. This is something that is in use in most games today on some level or another. However, there were some distinct points where it was very noticeable that the AI was switching. For example, once you crossed the finish line, your team either went into “celebrate” mode or “whiny pout” mode. These were simply repeating animations that continued throughout the duration of the post-race promenade (for example, while you are waiting for your wife to finish because you left an ocean of banana peels for her to wade through causing her to look strikingly like Kristi Yamaguchi doing non-stop toe loops down the final stretch). However, if you are to get whacked in the behind by an errant shell, you will immediately transition into the crash state, then the “stunned recovery” state for a few cycles of the animation (about 2 seconds). Then, magically, you will begin to celebrate once again. It makes it very easy to see that there are very defined edges to the states (and likewise the animations). The major states, then, seem to be as follows:
There are some sub-states mixed in there for the more subtle animations, but they are usually very short-lived. The driving mode, especially, has extra states for turning, sliding, punching, taunting or even just looking around because there’s nothing going on.
The AI racers don’t necessarily navigate freely along the track. They follow the technique of using invisible paths laid over (under?) the course. There are multiple lines on any given track and the AI will choose which one to follow at any one time. ![]() The reason that this became apparent to me was the first time that I saw two opponents in front of me fighting for the exact same stretch of road. There was more than enough room for them to maneuver, but they insisted on being in the same spot for some reason. What’s more, that “spot” was meandering back and forth along the course… as one car drifted left, the other went right along with him. As they drifted back, I’m not entirely sure what causes the AI to switch paths. Obviously an event like a spin or crash would cause the AI to pick up a new line. However, it is difficult to tell of the AI changes lines when nothing else is happening – like a cross-country skier stepping into a new pair of tracks. This does, happen, however. I’m just curious as to when and why. Certainly not when they are bumping and grinding down the straightaway, however. There is another factor, it seems, behind why the paths are the way they are. Not all paths are created for all cars. If the AI is in one of the big heavy cars, for example, they will tend to pick a path that wanders back and forth across the track. This is very noticeable with That brings me to another point… rubber banding. The Search For Equality…
First there is the speed of the cars. In general, when you start a new “Cup” (group of races) and the game “picks teams”, the AI will also assign a general skill level for each car. Throughout that cup, the running order will generally be the same give or take a couple of places. The same 2 cars will generally be in the top 3 and the same two cars will generally be in the back of the pack. This is obvious after having finished a 16-race cup and having one car have a single-digit score – which is only possible if you are continually finishing either last or 7th. There seems to be three ways they go about stringing out the pack like this: 1) By simply using the published speed of the cars. The fast cars will tend to be toward the front, whereas the putt-putt-mobiles will be in the rear. That means that those first two cars you have to contend with as a good driver will be the big, fast cars.
Now, there does seem to be a change in that algorithm as the race progresses. If you are in last on the first lap, there isn’t as much of a sense of urgency. If you are in first on the first lap, you won’t be getting assaulted non-stop. However, in both of those positions, things seem to change as the race wears on. If it has been smooth sailing for the first two laps, you can expect to have attracted a bit of attention on the third one. Perhaps it is this phenomenon (and the resultant indignation that ensues) that causes people to cry foul. However, once you have been in the position of playing catch-up on the final lap, you will realize that the game is more than happy to dispense one killer weapon after another in order to help facilitate retribution upon the AI. This is even in play on a multi-player race without the AI even being present. So again, this is a game design decision in order to further exciting races. In the Pursuit of High Drama… A couple of notes on artificially generated excitement… I can’t tell you how many times (probably because it is now into 4 digits) I have been on the final lap, in 2nd place and within striking distance of the 1st place car, but seemingly without hope… Another observation I have made in this vein is that things that green shells don’t seem to have studied for their Physics 101 final. I swear that green shells, upon hitting walls or other obstacles, will uncannily bounce so as to cross in front of you. It happens too often to be coincidental. Similarly, I have seen green shells bounce off of objects at unlikely angles to strike the AI car in front of me a la the “lone banana” phenomenon above. I know that this is not a hard computation to make… just a little quick vector math based on the path of your car. Therefore, it is my estimation that this is a subtle trick to make the game environment a little busier in the right way. Make it look like things were really dangerous – especially when my back-seat sister forgets what she’s doing and throws green shells forwards instead of backwards. Things That Work…
The characters use their weapons well and intelligently for the most part. The driving makes sense, even when it wanders around the track. The character animations are engaging and amusing at times. (Being a Yoshi player, I am always tickled when he stands on the back of the kart looking “vigilant” glancing left and right to see what’s up.) Things That Could Improve… Sometimes, I wish that the AI states were just a little more obscure. It isn’t distracting or annoying, but I want there to be a little more subtlety to them. This is not that big of a deal, however, especially since we aren’t dealing with life-like, human characters. The effect goes quite well with the cartoony feel that is a trademark of the Mario games.
So… is there anything that you have experienced that I may have missed? Is there anything I said in here that is completely out of whack? Drop in a comment! One thing I would be very amused to find out is that I have chalked some effect up to the AI when it wasn’t really there in the first place. Now THAT is the sign of a good design (or an AI programmer who wants to believe in things waaayyy too badly!). |
|