Oddball_E8 Posted October 23, 2003 Share Posted October 23, 2003 As we all know the AI in CMBB is... well i dont want to use the word flawed since its one of the darned best AIs ive seen in a computer game, but it isnt up to human vs human standard. fx it doesnt use command very well and usually send its HQs charging ahead of its "cannonfodder". it doesnt use tank formations of any kind and occationally has the idea of using unarmed and unarmored vehicles in the role of a tank. it never uses transports very well. well the list could go on and on... but recently ive discovered that it does simulate one thing very well... namely the disorganised russian forces at the beginning of the war... so if youre playing as axis in the early days (41-42) go ahead and play against the AI. but if you are playing as anything else you´d better find a human opponent... 0 Quote Link to comment Share on other sites More sharing options...
MikeyD Posted October 23, 2003 Share Posted October 23, 2003 What the AI needs is the guiding hand of an experienced scenario designer. someone to position the mgs and the mines logically, someone to place the tanks along a preferred travel route. Someone to drop in the right types of reinforcements at the right moment in the right area. Playing the AI against generic QB maps is a hit-and-miss affair (though I've had a few that were a jolly romp!), but playing a scenario (the first time through) that's been crafted specifically for AI play can be quite a challenge too. 0 Quote Link to comment Share on other sites More sharing options...
Oddball_E8 Posted October 23, 2003 Author Share Posted October 23, 2003 yeah i understand what youre saying, but i mean AFTER setup... there is no way to make the AI act in a coordinated way... it will send its tanks in the fray piecemeal... it will keep its units out of cohesion... it will make slow advances... no amount of scenario design is going to change that... 0 Quote Link to comment Share on other sites More sharing options...
Moon Posted October 23, 2003 Share Posted October 23, 2003 No amount of coding either I'm afraid. Not for the time being. It's not really breaking news that artificial intelligence isn't up to par with human intelligence these days A fully dynamic non-scripted AI like that in CM capable of playing any randomly generated map and force mix has strong points and weak points. I've had some great battles against the AI, especially if it's in the defensive but not only, and I've had some lame ones. But the point is this: there are AI's out there where a tank has trouble finding a straight route between two points; and there is an AI out there which is not up to par with an experienced human player in coordinating combined arms operations (and to be honest, I know several humans who'd do a worse job!). We're talking entirely different levels of AI quality here, just to make sure that nobody forgets this. Posts like these when quoted out of context have the nasty habit of creating false reputations... Martin 0 Quote Link to comment Share on other sites More sharing options...
carlrender Posted October 24, 2003 Share Posted October 24, 2003 Well I've been beaten by the AI (in scenarios) often enough to think it does a pretty good job. And there is absolutely no doubt it is -usually-FAR better than the AI all other games that I have played. 0 Quote Link to comment Share on other sites More sharing options...
Oddball_E8 Posted October 24, 2003 Author Share Posted October 24, 2003 im not complaining about the AI (look closely at my first post), im just saying that since the ai still isnt as good as a human, the best way of using it (IMHO) is as disorganised russians in the early part of the war... 0 Quote Link to comment Share on other sites More sharing options...
Sergei Posted October 24, 2003 Share Posted October 24, 2003 I doubt it really portrays Russians well... that Russians often wasted their strength and were unable to properly respond to rapid tactical developments, doesn't mean that when they had time to prepare an attack, that attack had at least some essentials of tactics in it. You're better off letting the AI defend, with a bonus. But human opponents are always preferable. 0 Quote Link to comment Share on other sites More sharing options...
SgtMuhammed Posted October 24, 2003 Share Posted October 24, 2003 MikeyD has a good point about the need to craft scenarios for the AI. One of the things I really hope to see in CMX2 is better control over the AI for the designers. I prefer to play and design operations and really wish there was more flexibility in the system. I know that trying to give the AI a real tactical sense is a nearly impossible programming task (for now) but it should be possible to define general behavior, (defend, attack, delay, agressive, passive, etc.) while still letting the AI control the actual unit actions. The current scenario designers work wonders with what they have but I think with a tweek or two it would be possible to work wonders. 0 Quote Link to comment Share on other sites More sharing options...
eichenbaum Posted October 24, 2003 Share Posted October 24, 2003 Also agree with MikeD. The placements of the units for AI is very important for it's response in the battle. And the testing before and after the design of a scenario is necessary to get a grip on the computer. It's to bad that the scenario designer is not able to set out the waypoints and actions for the AI's units in the scenario editor. Just do the set-up and the 1st turn. No scripting needed. 0 Quote Link to comment Share on other sites More sharing options...
Joachim Posted October 24, 2003 Share Posted October 24, 2003 Hey, 8 posts already and I second each! The AI is the best AI (for its genre). But the best sometimes just ain't good enough. Most of my games are GE vs AI soviets (QBs). I play them along BCR rules, usually the AI has a bonus. A list of some scenarios where the AI does a good job: a) setup for defense on custom-made maps with medium cover (ie defendable terrain). Placement of units is not outstanding, but solid work. setup for defense on maps with good cover - mainly because you just don't know where the AI will put his troops and can't bomb likely positions. c) Human wave infantry attacks with lots of higher HQs and a bonus to the AI. Any waste of ammo by the humn player is punished. Abundant higher HQs help the AI in maintaining some command structure. d) Dynamic flags with AI attacker!!! This is a real gem! The AI does not attack the nearest but the only valuable flag. It is much harder to outguess the AI then. The dynamic flags are often overlooked in custom scenarios! Any battle with a very huge AI bonus Things that don't work. Some of them might be able to change. AI defense: If the flags are too far from the AI map edge, AT gun placement is bad - especially if they can't even see the areas around the flag. AI defense: Counterattacks. Piecemeal, too early. Maybe tweak it so the first captured flag does not always trigger the counter. "Capture 1 outlying flag ASAP, wait for counter, slaughter counter, advance further" is faster than advancing into the counter. Sometimes you can just slaughter the counter and win without any other flag. AI attack: I'd like some more probing by the AI to see which flags are defended I'd like to stop the AI from taking the flags in a fixed pattern (the nearest flag to the start line, then the second nearest flag to the start line - even if it is on the other flank. Parading parallel to the front line is a recipe for disaster. The original plan is probably to roll up the flanks, but if the human player abandoned the forward positions, the AI is doomed. Penetrate deeper, then roll up the flank! Gruß Joachim [ October 24, 2003, 06:17 AM: Message edited by: Scarhead ] 0 Quote Link to comment Share on other sites More sharing options...
undead reindeer cavalry Posted October 24, 2003 Share Posted October 24, 2003 out of curiosity: does someone know a better AI what comes to this general style of setting? doesn't need to be WW2, doesn't even need to be tanks and soldiers. just something on the same scale etc. i don't know a single one that would come even close. 0 Quote Link to comment Share on other sites More sharing options...
Cookie Monster Posted October 24, 2003 Share Posted October 24, 2003 Well still being a newbie to this game I have to say I'm impressed with the AI. So far I've only played one scenario, Winter Wonderland, and I haven't been able to beat the Russian AI defenders yet! From what I've seen so far either the AI defends well or I just am not a great attacker. 0 Quote Link to comment Share on other sites More sharing options...
Hans Posted October 24, 2003 Share Posted October 24, 2003 One interesting thing I've noticed in trying to get a good attack out of the AI is to trick it into doing what I want. To me it seems that its movement towards an objective is often more energetic than the actual attack on the objective. I therefore try to lure a human player to move towards an objective knowing the AIs attack path to another objective will cause a collision. Not easy to do but great when it works! 0 Quote Link to comment Share on other sites More sharing options...
JasonC Posted October 24, 2003 Share Posted October 24, 2003 The only two things I've ever found a challenge with the AI are (1) when it attacks with +100% forces, thus 3:1 point odds. Believe me that is tough even with how dumb it is. The issue is ammo discipline and exchange losses. You have to kill everything efficiently and cleanly, or it just wears your men out before you stop them all. You can learn a lot about constructing a clever defense this way, and about which weapons are the best counters to various items. Sure, some things will be silly AI tricks (infantry AT ambushes), but most are real tactics that work against humans, too. And (2) when the AI has numerous bits of uber-armor, meaning stuff that can kill my thickest AFV, and at least sometimes get bounces from its own front armor facing against most of my AT weapons. That teaches tank driving, use of cover, teamwork, etc. This one is especially difficult when you also have to attack. 0 Quote Link to comment Share on other sites More sharing options...
Mailer Posted October 25, 2003 Share Posted October 25, 2003 if any of you have ever played the excellent game, "Steel Beasts" the scenerio designer allows you to easily script very complicated AI plans. You can use conditions to have them react in very complex or even random ways if you choose. Something like that for CM would be great. 0 Quote Link to comment Share on other sites More sharing options...
Hans Posted October 25, 2003 Share Posted October 25, 2003 CMx2 needs a system of hidden objective "wayposts" for the AI to react to. This would allow the designer to create more innovative ways to attack the player/foil his attacks. Additionally you need to be able to tell individual units what their "state" is. This would range from fanatical attack or DIP defense to engage and withdrawn or counter-attack after 'x' is taken. 0 Quote Link to comment Share on other sites More sharing options...
Sirocco Posted October 25, 2003 Share Posted October 25, 2003 There are some basic AI improvements that are well within the capabilities of the designers now, but conducting a strong attack in particular is a difficult task for an AI. The AI must be able to analayse the map, understand the strengths and weaknesses of it's own forces to create a plan, and then react to events, for example when to commit it's reserve. There are an endless number of variables to consider which makes AI design a huge and fascinating challenge. 0 Quote Link to comment Share on other sites More sharing options...
JasonC Posted October 25, 2003 Share Posted October 25, 2003 The AI could use a lot of improvement. What it needs, though, is a modest number of tactical dos and don'ts built in. It doesn't need to slice bread. But it should know things it doesn't. 1. With infantry, proper spread is more important than getting to cover. Right now it'll pack a company into a patch of scattered trees the size of a postage stamp. One HE round and they are history. It should estimate the size of unit that will fit - with tactical effectiveness - in cover along a route. If it has to send more use lesser cover or the open. In shorthand, don't overstack. 2. Avoiding obstacles is more important than cover or going straight to an objective. It bunches a column of men right over wire or through mines when 60m to either side there is nothing. Why? Direct at a flag, a narrow tree area with open on sides, etc. In shorthand, go around. 3. Avoiding kill sacks is more important with armor than getting LOS. Don't send every tank to the exact same spot. Avoid locations where tanks just as thick or thicker have already gotten killed. Platoon formations to get many on fews are good, but require some spread - at least 40m from any - live or dead - friendly AFV. Elementary traffic control. 4. Support weapons aren't infantry. They fight from range. Get to cover, then ask if they have LOS to any enemy. If they do, sit and shoot. Light mortars and HMGs don't charge flags. If they have no LOS to any enemy, head (at "move") toward a flag or a known enemy, one piece of cover at a time. Stop as soon as there is LOS. Similarly, HQs don't lead "by example". If nobody is following, stop charging. Pause if necessary, but keep some troops in command. 5. Defenders don't charge out of cover and prepared positions looking for the enemy. Unlike attackers, the preferred route for any relocation is a lateral shift with a rearward loop (step back, step right right right, step forward). It is fine to orient on seen enemy or captured flags, but don't charge the attack out of mere inability to sit still. 6. Call for arty with reactive modules (time to shoot 5 minutes or less). Not smoke on own positions, HE on enemy positions. Spotted enemy on a flag are a good target - shooting at friendlies best. A target is any set of enemy "markers", present or recent, but more than one. Fire for any fraction of a minute plus a full minute - not "a whole module on one gun" or without any adjustment or not at all. Long delay modules (over 5 minutes) should use planned fire at large bodies of cover. Near flags is better than not, but any cover inside the enemy set up zone is OK. Start is OK, but at 5 minutes or 10 minutes is also good. If it misses it misses. But use it, take your shots. 7. Basic movement drills. Advance over open ground in 50-100m bounds. Wait if "tired". Before action, "move to contact" in cover only. "Move" after contact. "Move" with all support weapons, or in mud or snow. Do not "sneak" more than 10m. AFVs fast move if not in effective range of known opponents, hunt if they are. If being shot at, first priority is to shoot back, with that unit and with others. Getting to better cover a distant second. 8. Send infantry first if you've got any. Armor does not approach covered areas within 50m unless friendly infantry is there already. With infantry, one unit per platoon half a minute ahead of the rest would be a nice touch. Elementary scouting and combined arms. Right now it drives otherwise unkillable tanks into infantry AT ambushes recklessly. These are the biggest problems the AI has. It should not be rocket science to add a few dos and don'ts. And just avoiding these constantly repeated, predictable, boneheaded errors - even 2/3rds of the time - would make it a much tougher opponent than it is. 0 Quote Link to comment Share on other sites More sharing options...
CombinedArms Posted October 25, 2003 Share Posted October 25, 2003 JasonC has a good list here. I guess the question is--how hard/easy would it be to code in some of these rules of thumb into an engine's AI. Human beings tend to operate by such rules of thumb--We tend to say, "Hmm, in these circumstances it's a good idea to do THIS but not a good idea to do THAT." Tactical ability is partly knowing the rules of thumb and partly recognizing when, where, and how they apply. It gets interesting when the rules thumb come into conflict and you have to make a choice. But my impression is that computers have a limited ability to operate according to such rules. Would it be possible, say, to code a game engine to recognize that its a bad idea for defenders to charge out of defensive positions around a large flag to protect a small flag that's clearly already lost. That one change would significantly improve the AI's defensive skills. Obviously, this would be for the next generation of CM. 0 Quote Link to comment Share on other sites More sharing options...
JasonC Posted October 26, 2003 Share Posted October 26, 2003 The problem AIs tend to have is selecting a rule to apply by recognizing which situation they are in, or the relevant context. But the rules I am suggesting should not have that problem. To heck with conflicts, you can just always do them. Will this occasionally be less than optimal? Sure. What does optimal have to do with anything? It won't be a tenth as boneheaded as what it does now. Computers are routinely programmed to apply such rules, once humans have spotted and formulated them. Chess programs are full of them. And they have much less room for improvement and thus tolerance for less than optimum rules of thumb than a CM AI has. You can formulate these things in pseudo code easily enough. For example, say you want to tell the AI what to do with any arty modules it has. If attacking and delay for the module greater than five minutes, then fireplan with that module. If attacking and delay five minutes or less, fire plan if rockets. Else fire plan with single module with largest blast rating. All others, no fire plan. Will this be a perfect division of arty use? No, just way better than what it does now. Also relatively unpredictable. If fire plan attacking, 35% first turn, 10% each 2-5, 5% each turns 6-10. Will it time everything perfectly? No, but defenders will get hit, movements will sometimes be restricted by falling shells, occasionally a barrage will come right ahead of an infantry attack, etc. Where to aim it? Not "only if you see a truck on a hillside at set up". Aim at the largest bodies of cover in the enemy set up zone. Aim at bodies of cover near the flags. If in doubt, aim "target wide" at the middle of the defender's zone - it is better than not firing, or blowing all arty ammo on turn 16 against one gun on the left flank. Reactive fire can also be addressed by pseudo code rules of thumb. Count points, and when you get to "4" within an 80x40 box, order a fire mission. As follows - Nationality flag - 1 Any vehicle - 1 Any towed gun - 2 Spotted infantry - 2 Shooting at friendlies - +1 Then let the battery count down. When it reads 2 minutes, adjust the point of aim to maximize the number of points under an 80 long, 40 wide box if doing so will increase the "count" inside the current box by 50%. When only 1 minute remaining, only shift it doing so will double the "count" inside the box. Will this be perfect? No. But it won't fire at a single unit, it will do some adjusting, it will put arty on bodies of cover whole companies of infantry have accumulated in, etc. When a defenders gets 4 "points" under a TRP the time will be under a minute immediately so it will just put rounds there without more ado. This truly isn't rocket science, people. 0 Quote Link to comment Share on other sites More sharing options...
Sirocco Posted October 26, 2003 Share Posted October 26, 2003 When considering AI, you have to appreciate that changes that seem perfectly simple can have unforseen consequences. An interesting example is the defender leaving foxholes to counterattack. It often does so unwisely, but the intentions behind it are good. The addition of memory to units will make a huge difference to the AI, right across the board. [ October 26, 2003, 07:40 AM: Message edited by: Sirocco ] 0 Quote Link to comment Share on other sites More sharing options...
Battlefront.com Posted October 27, 2003 Share Posted October 27, 2003 If AI was so easy to program, as Jason claims it to be, dontcha think game AI would be better all around? Unfortunately Jason's concepts are naive in terms of the difficulty of implementing easy to grasp concepts into computer code. CM's AI could be much better, for sure, but it could also require 500MB of more RAM and take 15 minutes to compile turns. As is often the case with amateur advice, it is just that Steve 0 Quote Link to comment Share on other sites More sharing options...
Oddball_E8 Posted October 27, 2003 Author Share Posted October 27, 2003 Well ive voiced my oppinion... this was never meant as a discussion about how to change the AI. i just wanted to express my views about what the AI was best used as. 0 Quote Link to comment Share on other sites More sharing options...
JasonC Posted October 27, 2003 Share Posted October 27, 2003 Horsefeathers. It is not processing or program size (modern desktops can run a Mars mission, easily), it is just time spent programming it up, debugging, and refining routines. It is the input time that takes work, not the number crunching. 0 Quote Link to comment Share on other sites More sharing options...
Joachim Posted October 27, 2003 Share Posted October 27, 2003 Originally posted by JasonC: Horsefeathers. It is not processing or program size (modern desktops can run a Mars mission, easily), it is just time spent programming it up, debugging, and refining routines. It is the input time that takes work, not the number crunching. Depending on the size of the enemy forces, the time to compute the AI turns is very long. I guess the problem with the AI is that each unit plans its own mission, unaware of the others. Resulting in n calculations. If you plan it calculating what all the other units are doing, you will get from n calcs per turn to n*n calcs. So with 100 units, you will spend on the order of 100 times longer than now. If you use formations and sub-formations for movement and do the calcs based on them, you might get less calcs, but I guess it would still be a lot of calcs more (say you have 25 plts and 100 unit´s. Guestimate is 25*25+100 calcs). I have done some work in optimal control of stochastic dynamic models (ie trying to get the best result over several time periods (aka turns) where you don't know the exact behaviour of things and have some added uncertainty). There are some approaches now adding game theory (ie another player). But then optimization is even more complicated. The AI behaviour JasonC wants (especially considering for one unit what happens to the others) could be modelled that way. I guess "Bit Battles" would be the favourites of 99,9% of players then - all others would take hours to compute 1 turn. If AI programming was so far ahead, it could solve our problems, too. Then we could stop research... BTW: I want that good AI, too. But I believe BFC when they say it will take lots of computational time. Gruß Joachim 0 Quote Link to comment Share on other sites More sharing options...
Recommended Posts
Join the conversation
You can post now and register later. If you have an account, sign in now to post with your account.