StarCraft AI Competition Results

bgweber writes "The StarCraft AI Competition announced last year has come to a conclusion. The competition received 28 bot submissions from universities and teams all over the world. The winner of the competition was UC Berkeley's submission, which executed a novel mutalisk micromanagement strategy. During the conference, a man versus machine exhibition match was held between the top ranking bot and a former World Cyber Games competitor. While the expert player was capable of defeating the best bot, less experienced players were not as successful. Complete results, bot releases, and replays are available at the competition website."

Comments

I for one

[Anonymous Coward]

am terrified we're actively pursuing the ability for our robot overlords to have the perfect micromanagement strategies.

Re:I for one

[durrr]

Overlords can be spawned from hatcheries, your real enemy is the robotic overmind. Know the difference, it could save your life.

Re:I for one

[darthdavid]

Yes but that's mounted on a warship, that mode can be turned off (by manually disconnecting the weapons if necessary, there are people on board after all) and that mode would only be turned on if a shooting war broke out and they needed to make sure that the ship wasn't going to get blown up before they could hit the "shoot down that motherfucking missile screaming in at us" button.

It all adds up to something a bit different than an autonomous weapons platform that can choose to engage targets at will, which is what other people were talking about.

Not equal

[Pojut]

By default, I'd say that a well-designed bot in an RTS would have an advantage over all but the best players. Since there is so much to keep track of, software would win out over wetware.

Bots in an FPS are one thing, but when you have dozens of units, a build order, multiple fronts, resources, and more to track all at the same time, the infinitely scalable multitasking of a bot would certainly come in handy.

Or maybe I'm off base?

Re:Not equal

[Alsn]

A bot would always be superior in multitasking, but the most important ability in Starcraft(which the article is about) is decision making which is really hard to program for an RTS bot. Or at least, so it would seem as I've never seen an RTS AI that hasn't cheated and at the same time been challenging.

It doesn't matter if you have the best multitasking in the world if the opponent can just outright kill you cause it has more stuff.

Re:Not equal

[imgod2u]

It's a huge advantage but it's only one part of the game. Many would argue that overall strategy is far more important than micro-management.

One of the top protoss players (WhiteRa) in SC2 (former SC1 player) has a relatively low actions-per-minute (APM) count compared to most players. Yet he still comes out on top by a lot.

Being able to multitask and micro-manage is definitely an advantage but a far more important ability is being able to plan on the larger scale. I've never seen an AI capable of harassment techniques, guerrilla warfare and exercising map control. Multi-pronged attacks are also something that it should theoretically be great at but it never really tries.

Re:Not equal

[Monkeedude1212]

Bots in FPS ARE a completely different thing, as setting up an auto-headshot sniper that will hit the player before he can get the bot into view is not that difficult to make. That field will never be an even battle.

Strategy games are a little different though. A Bot essentially has no "better" way to evaluate the player than any other player would evaluate the player. Say in Starcraft, the Bot scouts the player - and determines he is a little behind in what he would expect the player's army count to be. This could mean a number of things: The player made a mistake, the player is saving up, or theres something the player has that you have not found yet. How do you proceed?

Now - when you get to the pro level of gaming, you worry a little less about your opponent's build and worry more about not letting them know yours. Walling and other defensive techniques become just as important as scouting your opponent. The game becomes highly a higly reactive scenario as opposed to proactive. If you know what your opponent is doing, you can counter it and that puts you much further ahead, possibly ahead enough to crush them.

So the problem eventually lies in getting an AI to properly counter a players actions. Making an AI react to players is much harder than giving an AI a plan and telling him to execute. Because essentially the reaction is only as smart as whoever is programming the AI. And if you are a better player, capable of keeping other people from determining your plan, you can beat an AI who is trying to determine yours.

Don't get me wrong, the ability for computers to instantly Micro and Macromanage all of the units and resources at once does give it some serious advantages, but deep in the heart of it: The AI will only be a little better of a player than the person who programs him. (Or her, if you program female AI's like GLaDOS)

Re:Not equal

[bytestorm]

Ideally, you can watch your opponents prior matches and tailor your strategy to defeating him. A computer might have a harder time doing that, but a person would be able to recognize the weaknesses of a strategy on the spot. Starcraft is very much rock-paper-scissors.

Re:Not equal

[e4g4]

Starcraft is only rock-paper-scissors if you don't scout.

Re:Not equal

[immakiku]

Examples like that are rare. Taking your example of SC2, there's few scenarios where superior micromanagement can turn a big disadvantage. From experience, most silver+ level players do not just NOT micromanage their units. They just do it without taking a hit in their macromanagement or they do it less effectively than top level players. Even in the reaper vs ling/roach scenario, a player who's seen the situation more than once will know how to properly react (without needing to excessively waste APM defending against the reapers).

Re:Not equal

[Alsn]

Of course it can make a huge difference, but no matter how good your multitasking is, tanks still don't shoot up...

The point is, it's really easy to just outsmart an RTS AI once you figure out its tendencies. No matter how good it's multitasking is your own won't be bad enough that your flying units all of a sudden forget how to shoot down.

Having an intricate knowledge of counters and economy and timings(at least as far as starcraft goes) has always been superior to multitasking, but only to a point. There are things that you can take advantage of with near-perfect multitasking and control, but you need to actually know what to do with that control(and so far, no AIs do). A good example of what you can use AI for is in the article video demonstrating mutalisk/scourge control at a level which no human player would be able to accomplish. The only problem is that a human player would know that the AI is capable of that and just play around it.

An example would be some extremely mechanical starcraft players(NaDa, a korean progamer comes to mind) being able to just completely outdo their opponents with pure control and multitasking. The problem here though is that those players have such an immensely better understanding of the strategy and tactics of the game than a programmed AI can ever come close to even if it's not their strong suit(compared to other progamers who play more with smarts than raw hand speed). Their multitasking wasn't what won them games but rather a combination of their abilities. They still had far and away stronger decision making than your average player(or any kind of AI).

Re:Not equal

[kaizokuace]

I agree with you whole heartedly. I am a Go player, so I understand that strategy and intuition come in to play with this type of game where "war" is concerned. The most advanced Go programs can't defeat a low level player. Being able to look at the board as a whole and not a series of spaces seems to be beyond these AI. In Starcraft I would imagine that an AI can't plan ahead for the types of actions a human would take. Being able to micro manage and deal with skirmishes and build orders and such is just a mechanical action. Just the basics of combat. Placement, actions, and timing of such is what matters. Not just that you send out a group to attack but where and when matter in the whole of the match. Planning ahead, asserting influence, being creative. These are the abilities humans have over the AI.

Re:Not equal

[UnknownSoldier]

> No human can really keep up with 2 100% controlled and coordinated attacks. Mainly because we have to split our resources between them. The AI could realistically perfectly control 2 attacks with all the skill that a human player could devote to a single attack.

The biggest problem is the INTERFACE of the game. Let me know when I can create on-the-fly Picture-In-Picture overviews of the map in real-time, so I _actually_ can attack/defend on multiple fronts.

Sad to see RTSs really haven't changed in 20 years ;-(

Re:Not equal

[leonardluen]

as part of the rules for this contest the AI had to have fog of war turned on.

Re:Not equal

[Arancaytar]

Or her, if you program female AI's like GLaDOS

You will be Zergrushed. And then there will be cake.

Re:Not equal

[iluvcapra]

The biggest problem is the INTERFACE of the game. Let me know when I can create on-the-fly Picture-In-Picture overviews of the map in real-time, so I _actually_ can attack/defend on multiple fronts.

An interesting modification to StarCraft which would give AIs a run for their money would be cooperative play where several users operate one team on the field, and where the several users have a hierarchy and delegate command of corps of units to other players, maybe one player handling resources while another scouts while a third consolidates the offensive force.

Re:Not equal

[Anonymous Coward]

Not "would" - does. That feature has been in the original StarCraft ever since. The original Age of Empires had that feature, too.

It's sad that they didn't include this in SC2 though.

Re:Not equal

[aeroelastic]

An interesting modification to StarCraft which would give AIs a run for their money would be cooperative play where several users operate one team on the field, and where the several users have a hierarchy and delegate command of corps of units to other players, maybe one player handling resources while another scouts while a third consolidates the offensive force.

This is basically what we would do in 4 v 4 Age of Empires. The person farthest from the other team would be "resource bitch". They would focus on tech upgrades to mining and farming, and be responsible for constantly supplying the other players with food, stone, etc. The other players would have no economy but would build cavalry, siege weapons, etc based upon what tech their civilization was best at.

Re:Not equal

[MoriaOrc]

Just to add on to your point. In the Mutalisk micro video, it's pretty clear that the mutalisks are prioritizing destroying siege tanks over goliaths. While that might be all well and good in some sort of base defense, or ground/air simultaneous attack, the videos are all of an air-only attack on the enemy base. The siege tanks are literally doing nothing but taking hits, while the goliaths represent the only actual threat to the Mutalisks (except the turrets in the final part of the video). Add on to that that IIRC goliaths take less damage to destroy then siege tanks.

Most human players would realize to attack goliaths before siege tanks in the same scenario, since they are both a softer target and a more immediate threat. Destroying them first means the Mutalisks could continue to harass the enemy base without taking any harassment back. However, the AI doesn't realize this and keeps attacking siege tanks whenever they are present.

Important note

[Monkeedude1212]

Uses Broodwar, not Starcraft 2, and not just the original.

Just saying. (Cause when they mentioned Ultralisk Microing, I thought about SC2 and how Ultralisks are terrible units there simply because they block your units making Micro a huge pain, and it wasn't so bad in BW when your units could take a bit more of a beating).

That's the best strategy?

[jandrese]

So the winner is just a Muta harassment bot? I have to wonder if the top level human player just spammed some Corsairs or Valks? It didn't seem like the AI was particularly good at changing strategy if the opponent countered. Spreading out the Mutas would help against a Corsair or Valk counter though, since both of those units rely on the Mutas natural tendency to clump to get maximum effectiveness.

Re:That's the best strategy?

[Anonymous Coward]

Against Muta's? Just spam Goliaths.

Re:That's the best strategy?

[jandrese]

The bot looked like it was pretty good at dancing the mutas around cliffs though, which is usually the downfall of Goliath spam.

Re:That's the best strategy?

[teko_teko]

It's actually pretty interesting technique that they used. They don't just clump up the mutas. From the results page:

Contemporary StarCraft wisdom tells us that the best way to use mutalisks is to clump them. In human versus human battles, this makes it difficult to single out the weaker mutalisks, because the units are stacked on top of each other. However, UC Berkeley’s team identified a flaw in this tactic; it reduces the damage output of each individual mutalisk, because not all mutalisks will fire when using this tactic. Instead, they employed a model in which mutalisk are always moving, maximizing damage output while simultaneously maximizing movement.

Video can be found at the bottom of the page.

Re:That's the best strategy?

[dburkett]

I'm a member of the team that built the winning bot (the one that wrote most of the mutalisk code, actually). One of the things we found out is that, as an AI, it's completely possible to keep the mutas' dps up without clumping them -- humans mainly have to clump them because they can't issue individual orders every second. When the bot is up against corsairs or valkyries (or any other unit that does splash damage), the mutas deliberately spread out to avoid getting splashed. Corsairs are still a good counter, but our mutas can hold their own against corsairs much better than most players. To address your bigger question, the choice of mutas was deliberate. Building any sort of unit controller is just hard and time consuming, so time constraints forced us to specialize, but we think we did a pretty good job. See another team member's post called "Overmind agent overview" for more details.

Re:my 2 cents

[grimJester]

When you play a game yourself, you only have to come up with a counter strategy to what your opponent is doing. When you program a bot, you have to think of all the possible strategies to counter and find reliable ways to recognize them in addition to having the counter.

Any 'learning' bots?

[JSBiff]

Anyone know if anyone has ever created a bot that has the ability to learn from losses and wins, to figure out what works and what doesn't (both what it is doing and what the enemy is doing, then use that in the future to predict what the other player is *trying* to do and come up with a counter)?

I've not played a lot of RTS's, but I've played a few, and the thing I've noticed is, if a strategy works against a bot, even if there's a fairly obvious counter, it will always work against that bot.

The RTS I've played the most is Supreme Commander: Forged Alliance, so I'll use an example from that: There are a collection of different bots available which will use different strategies. Now, the player can build a strategic missile launcher, which builds nukes. You can also build a strategic missile defense which will build 'seeker' missiles which will shoot down incoming nukes.

Some of the bots will build SMD's, some won't, but in no case does it build an SMD based upon the player/opponent's actions. That is, it doesn't check to see if you have started to build any Strategic Missile Launchers before it begins to build the defenses. The bots that don't build defenses won't build them even if you are building one. This means on the one hand that the bot wastes resources which it could have used otherwise, to build defenses, while on the other hand, if it doesn't build them, you can pretty easily and quickly defeat the bot with a nuke or two. Alternatively, instead of building a strategic missile defense, the bot *could* try to use someattack method (for example, if you don't have good air defenses, it could hypothetically try a targeted attack with a bunch of bombers or gunships to either destroy the SML, or the engineer units which are constructing the SML).

However, the bots never seem to be smart enough to attack the obvious threat of a strategic missile launcher. It seems like the only way the developers found to make the bots harder is to make them much more efficient at building up their economy and spamming out lots of land-units to try to attack the player.

When I get a chance, I want to try SC2, but right now, I'm in a period where I'm not playing games as much as I used to, and trying to reduce my gaming down to almost none while I get some more important things done in RL. It'll be interesting to see how the AI differs in that game.

Re:Any 'learning' bots?

[Dakman]

A bot that learns from loses? An interesting game. It seems the only winning move is to be Korean.

Nerd Fantasy Extrodinaire: Ingame Scripting Agent

[Wrath0fb0b]

Competitions like these are great, but they (and games like SCII) really make me pine for an RTS with a robust in-game scripting language. I would like to write complex auto-executing instructions for my units like "Pursue but never enter the firing zone of a known enemy turret or siege tank" or "If your energy > X and enemy of type Y in range R, cast spell S" or "If you are unit type X, always position yourself between friendly unit type Y and the enemy". You could confine the script to some reasonable specifications (say, no more than 1000 queries and 100 orders given per second) if you want to deter brute-force approaches.

This is a totally different problem than writing a good AI, you would be focused on writing powerful tools that aid, not replace, a human player by letting him specify his intent on a higher level than "go here" or "attack that". Better visualization of what's happening would be an integral part of this too. I would love to have SCII give me an overview of what units/building I have, what they are doing/queued to do -- even better if they are grouped into functional 'squadrons'. Being able to have multiple panes/monitors looking at different things would help to.

I guess what I'm saying is that I'm old and slow and want to leverage my experience writing rule-based logic to beat the whippersnappers that can click faster than I can and keep more things going in their heads at once :-P.

Several non-obvious flaws

[Myji Humoz]

There seems to be a very wide range in the abilities of the winner and runner up bots that might not be noticed by someone unfamiliar with Starcraft. In order of appearance:

1) In the flying units versus flying units match, mutalisks (guys with wings) should have focused the scourges (little 'c' shaped guys) because scourges have about 1/5th the health of a mutalisk, but can suicide to take out 5/6ths of a mutalisk's health. Red ai focused scourges, the other ai didn't, with disastrous results for the other ai.

2) In the match with infantry, the players had medics, which heal other units but can't attack, and marines/ghosts, which can do damage but can't heal each other. One ai moved medics with ghosts such that the medics could actually heal. The other ai just left the medics a mile away from the combat, and got slaughtered. Furthermore, the AIs didn't bother with formations, which meant that half their units spent the entire fight trying to get into range. A precombat formation lets almost all the combat units start firing as soon as the fight begins.

3) The red zealots retreated in the face of numerically superior opponents, while the teal zealots just attack moved no matter how many they had. Teal zealots didn't focus fire, which meant that they lose units sooner, and thus had less damage output compared to red. In addition, red failed to kill the pylons (600hp) powering the buildings (a lot more than 600hp). Neither player built their bases to maximize the number of pylons powering their vital gateways; each pylon usually powered only one building.

4) In the fight with dragoons (orange spider things) versus tanks, the protoss (orange) could have frozen half the enemy tanks with a single stasis spell by sneaking the arbiter (flying spider thing) to the back of the tank formation. Furthermore, the protoss could have focused the science vessel (floating teal circle) that was preventing them from being invisible. It would have been a slaughter if the vessel had been focused, as teal would have had no real way to hit the orange units.

5) In the match between Overmind and Krasi0, the article talks about mutalisk clumping preventing some of the mutalisks from attacking. However, the point of the stack is that when one guy is in range, everyone else is. Also, the attack animation is so fast that for all practical purposes, the flyers can shoot while moving. They use their mobility to get out of range of infantry units, then zoom in to pick off outliers when their attack cooldown is finisihed. In actual competition, the terran player would usually rely on a strategic placement of static defenses and a highly mobile cluster. However, the terran (defending force) built tanks, vultures, and goliaths (mech guys that shot missiles) with the flaw that tanks and vultures can't shoot air, and are thus almost useless versus the mutalisks. The mutalisks spent the second half of the clip shooting tanks rather than focusing down the goliaths volleying missiles into them. The terran AI prioritized repairing tanks as much as goliaths, and didn't place tanks next to goliaths to soak up bounce damage from a mutalisk.

In general, the AI had problem with understanding the priority in a fight. That is to say, they often had no sense of what units are critical to a position or what units pose the most threat. They didn't arrange their units to maximize their effectiveness, and often failed to alter their behavior based on the other party. It's a fun contest, but I'm not sure the AIs could beat a moderately skilled player who understands tactics AND strategy.

Re:Nerd Fantasy Extrodinaire: Ingame Scripting Age

[ThePyro]

Have a look at the Spring Engine if you haven't already. There are a variety of RTS games, including some high-quality variants of Total Annihilation, which use the Spring Engine and allow for all sorts of client-side scripting through Lua. There are a variety of client-side lua "gadgets" that players have written already. You can move your units into custom formations by drawing lines or squiggles with the mouse; there are widgets to automate using air transports to ferry units between factories and rally points; there are even widgets to automatically alert the player when certain dangerous units are spotted. IIRC, someone was even working on a script for kiting with long-range units.

Generic VS Specific

[Nemyst]

I think quite a bit of this can be attributed on generic versus specific behavior. I am not an AI programmer, and I just glanced at the page and videos, but from what I gathered most of the AIs had hardcoded strategies that they would follow blindly. Then, the strength of the AI was directly proportional to the quality of the strategy that was implemented. They would use very specific patterns and follow them systematically.

What I would be curious about would be a more generic AI. Instead of using known strategies, the AI would analyze the battlefield and would, for instance, focus fire on support units (not on "medics" or "science vessels", but using the stats of the unit to determine its usage and purpose) or the most dangerous units (air-to-ground would prioritize fire on ground-to-air units) or the most important buildings (using ratios and comparisons to determine this, instead of hardcoding that building A is critical). Technically, it could then be able to react to a lot more situations and would also be adaptable if the fight changes or if the enemy bot has a counter for whatever strategy had been planned. Obviously, a generic bot is harder to build and it might also not be as efficient when compared to singular strategies, but it would be an interesting experiment to make.

UC Overmind

[dlwh]

I'm with the winning UC Berkeley Overmind team. ( http://overmind.cs.berkeley.edu/ ) We're very excited to have won the competition, and we're hopeful that soon we'll be able to create an agent that can beat the very best humans.

Here's an overview of our strategy. Broadly, our agent always plays Zerg, and its primary objective is to apply constant pressure to our opponents, which will let the agent continually expand to improve its economy. To that end, our agent had three primary tools: a scouting worker that harassed early, zerglings that provided early defense, and mutalisks which basically force the opponent to stay in their base. Our agent would make decisions, based on what it observed, to trade off between these different forces. It might build more zerglings if it sensed more early pressure, or it might skip them altogether if our opponent wasn't going to attack early.

Our choice of mutalisks was deliberate. They are a highly mobile all-purpose air unit that can mass up and cause significant damage in a short period of time. Also, they don't have bounding boxes, meaning that they're more amenable to computer control. That choice proved to be really valuable, because a lot of the opponents seemed to have preferred ground armies. However, there are many other strong units. For example, the other agent in the finals (Krasi0) was truly impressive with its ability to repair units.

Anyway, on our page we have a couple of videos, and we'll have several more by tomorrow (Saturday).