| Forum Index > BW General |
|
djsherman
United States140 Posts
I'd like to hear from the StarCraft community: what would you like to see from this years competition, as far as organization, results, and rules? Last years competition was a huge success and we are excited to see what future competitions will show us. UC Berkeley won last years main event with raging Mutalisks, what else can AI show us? | ||
|
ArbAttack
Canada198 Posts
| ||
|
djsherman
United States140 Posts
| ||
|
gen.Sun
United States539 Posts
A round where a 'surprise' element is added to the game on tournament day, in order to check for AI's adaptability. In mirror matchups maybe a unit is removed. For example, zerglings or mutas in ZvZ. Or maybe a special Python where the the main has only 4 mineral patches. A sudden decrease of CC/Hatch/Nexus price to 75/100/100. (ehh, probably imba in favor of zerg) These are some example surprises. This way we can see whether the AIs are truly intelligent or have hard coded BOs. From my understanding all the current bots would fail pretty hard if tested like this. But with another year maybe it's enough time to write something more flexible? | ||
|
djsherman
United States140 Posts
[B]This way we can see whether the AIs are truly intelligent or have hard coded BOs. BOs are expected, due to the complexity of the domain. But I do like some of the ideas you are proposing. Basically, there are two directions in which to increase complexity. The first is to relax the constraint of a fixed race, make the bot play all races, against all races, even random. The second is to relax the constraint of know maps, make the bot play on new maps. But AIs are not yet able to overcome humans when following both of these constrains, so why relax them at this point? StarCraft is an excellent domain for AI research and there are all sorts of potential for investigating general intelligence techniques, but first consider humans. Humans have to first study maps, or at least be given a high-level description of map features. Second, even top level players don't random races, because its too complicated, even for humans. I agree that an advanced AI would be capable of these feats, but it is far beyond the current capabilities of systems. | ||
|
gen.Sun
United States539 Posts
An AI should be given all the information that a newbie human is given, data on unit HP, attack, move speed, range, etc, and tasked with figuring out unit counters themselves. They should be able to figure out on the fly the answers to questions like, I'm expecting mutas, should I build more turrets or make barracks for more marines. Year 1 should be enough for figuring out all the mundane stuff like unit micro and pathfinding. Year 2 should be dumping the constraints for more interesting play. If the AI is thinking on a high enough level then things like race or map should not matter to them. If you set the rules from day 1 to favor flexibility then you might be able to encourage faster development in that direction. Otherwise teams might be tempted to stuff the bots with mid to low level strategies, like BOs and unit counters and end up with effective but uninteresting bots. my 2 cents. | ||
|
djsherman
United States140 Posts
If you set the rules from day 1 to favor flexibility then you might be able to encourage faster development in that direction. Otherwise teams might be tempted to stuff the bots with mid to low level strategies, like BOs and unit counters and end up with effective but uninteresting bots. my 2 cents. My view has always been that before we build general intelligence, we must first build specific intelligence, With chess, we went too far, and built chess specific machines, but as long as we are using cpus off-the-shelf we should be fine in our quest for intelligence. We can't build expert general intelligence until we have first built expert intelligence for a specific domain! | ||
|
kamikami
France1057 Posts
| ||
|
zobz
Canada2175 Posts
All i want to see is more full tournaments with lots of games played on proper maps, ala the final stage of last year's competition. For micro maps to be worth a damn they really should have dynamic armies of differing races, and then balancing the map would be sort of a problem and probably just a waste of resources, unless you have them to spare. What i really want to see is more stuff like that article on Berkley's Overmind. I find it extremely interesting to read about the whole process of game theory based programming, especially in relation to starcraft. | ||
|
stafu
Australia1196 Posts
 | ||
Spazer
Canada8031 Posts
In addition, more interesting unit compositions for micro maps would be nice instead of just plain mnm or dragoons. Something like tank/vulture/goliath/dropship. With a composition like this, you can do a lot of different things. Since this is essentially TvT, positioning becomes key. Where do you set up tanks? Can you lure the enemy within range of those tanks? How should I place the few mines I have? Can I effectively utilize my dropships to make their tanks splash friendly units? To date, the majority of the terran bots I've seen tend to be very sloppy regarding positioning. Oftentimes they'll siege up over their own mines, with predictable results. Stuff like that shouldn't happen. Basically, I want to see more emphasis on positional play. This is the one area I think bots really need to improve in. | ||
|
GGTeMpLaR
United States7226 Posts
1) No AI is capable of competing on a professional level in Starcraft we seemed to agree on this point but they claim that it is possible to program an AI that would be capable of outmicro-ing professional gamers rather easily while I was under the impression that this would be one of the hardest things to program am I mistaken or are they | ||
|
plutonian
United States6 Posts
| ||
|
Goragoth
New Zealand1065 Posts
As far as what I would like to see: a detailed write-up from each team, how they went about making their bots, design philosophies, etc... like the article on the Overmind but not just from the winner. Maybe some sort of documentary that could be put on YouTube would be cool too, with interviews with the developers and maybe some graphics to explain certain concepts etc... to make it all more accessible to a wider audience. Maybe someone could contact one of the science magazines like New Scientist to do a piece and maybe even a TV station like the National Geographic channel. I'm not sure what the funding and organization is like behind this competition but there are some fantastic PR opportunities here. | ||
|
rasdasd
United States82 Posts
In general, to what i have read on the previous posts. My bot, and pretty much almost all of the other bots, already have an algorithm to analyze the maps. My bot will spend the first 30 seconds analyzing the map if it has never played on it b4. As was stated before, The micro actions themselves are easy, but knowing when to micro is hard. I should probably post a video of my Psi Storm micro :D a 80 supply army beats 200/200 Hydras with 80 food of siege tanks supporting them :D I'll prolly post here again when i have more time. | ||
|
djsherman
United States140 Posts
On February 05 2011 18:13 Goragoth wrote: Why is it that every thread about AI invariably has some complete moron that thinks that any AI that doesn't have human-level intelligence is somehow worthless? We are at least 50 years from achieving that, probably more. Real, true intelligence requires levels of complexity that are hard to imagine. Having a computer solve even a constrained problem like playing Starcraft well requires a ton of work and interesting research. This is why we are starting with a fixed task. Not sure about your timeline of general intelligence, but yes, playing expert StarCraft is only the beginning in building human level AI systems. | ||
|
gen.Sun
United States539 Posts
| ||
|
rasdasd
United States82 Posts
| ||
|
nalgene
Canada2153 Posts
| ||
|
Ookm
Bolivia18 Posts
| ||
|
CoWsGoesMoo
250 Posts
| ||
|
Rlentless
Hong Kong322 Posts
http://eis.ucsc.edu/BotInstructions And there should be a link on the words "part 1" and "part 2". Of course read everything. Note: Most of these bots you need different versions to run them. | ||
|
antrax
Peru191 Posts
Did AI's consider worker micro for harvesting purposes? With constant micro AI could achieve maximum harvesting with 2 workers per mineral patch, hence faster macro mode. In 1vs1 this little difference could give one side real economical advantage. | ||
|
YejinYejin
United States1053 Posts
Mutalisk micro is simple enough for the AI to do. They just have each of the mutalisks do a moving shot, retreat, and then repeat. Vulture micro is simple, too. Move the vulture away, patrol click back at a specific angle, and then right click away again. How about the concrete micro? Like in PvP, if each player has 2 zealots and 3 dragoons. I remember it happened at some point in a Stork vs Bisu game on Benzene, I think, and Stork came out of the battle without losing a single unit. Bisu's were all gone, and Stork's were just heavily damaged, but every single one was alive somehow. Or in TvP when Terran does Fake Double, and Protoss fights it with 3-5 dragoons. This micro is not a single action that you can perform easily over and over again given 1000 apm. It's a constant decision-making process. | ||
|
Azzur
Australia6259 Posts
| ||
|
YejinYejin
United States1053 Posts
| ||
|
ReketSomething
United States6012 Posts
On February 05 2011 11:43 gen.Sun wrote: A round where a bot has to play all 3 races. A round where a 'surprise' element is added to the game on tournament day, in order to check for AI's adaptability. In mirror matchups maybe a unit is removed. For example, zerglings or mutas in ZvZ. Or maybe a special Python where the the main has only 4 mineral patches. A sudden decrease of CC/Hatch/Nexus price to 75/100/100. (ehh, probably imba in favor of zerg) These are some example surprises. This way we can see whether the AIs are truly intelligent or have hard coded BOs. From my understanding all the current bots would fail pretty hard if tested like this. But with another year maybe it's enough time to write something more flexible? I actually disagree with this stuff. Adding random elements isnt worth it. its hard enough coming up with a stable AI. adaptability doesnt matter at all....coming up with the best standard ai is better. | ||
|
Mumei
United States257 Posts
On February 07 2011 22:33 Azzur wrote: For a long time, people thought that computers could not compete with humans in chess. The best computers now destroy the humans players. I think the advantage the computer has in chess (brute-force calculation to move that gives them the most profit in the short-term (say, next 5 or 6 moves, the subtractive nature of the game, fairly easily to create an evaluative system) based on a system that gives a value to each piece (with the King naturally being given an absurdly disproportionate value to represent how important it is) won't really help in StarCraft, since decision-making is more complicated in SC than that, since it is somewhat harder to determine what the exact value is of a given choice - at least for a computer; human players tend to have an intuitive sense of what they should be doing. It's the same reason that computer go programs are still at about 6 - 7 stone handicaps behind top human professional players, which is fairly huge, considering how much of an advantage 6 - 7 stone is worth. As the Wikipedia entry for go handicaps puts it, "In points terms, one stone is considered to be 13-16 points, but this figure is not constant over levels: the more skillful a player, the greater the usefulness of each stone." In other words, it's basically a 80 - 110 point advantage; games of go should be won by less than 10 points if the opponents are evenly matched or the handicap is correct. Go programs, in other words, aren't anywhere close to beating top human players at even odds, let alone to the point where computer chess programs are (which can actually give odds to human players ala 19th century chess money matches). Computer StarCraft will have a huge advantage insofar as having the APM to do everything mechnically perfectly, but I think it'll have the same problems with evaluating positions and decision-making that computer go programs have, and the same inability to brute-force calculations for what the best move is (in chess, programs look a few turns ahead; this is impossible in go, because there are so many more legal moves in an average turn, and because of the aforementioned difficulty in evaluating positions). Perhaps the micro / macro perfection will allow the computer to overcome that weakness fairly quickly, but I'd really like to one day see a StarCraft AI that is capable of outplaying a person while playing merely peak human level, mechanically speaking - I think that'll be more meaningful. | ||
|
Trias
Netherlands53 Posts
On February 05 2011 10:51 djsherman wrote: I'd like to hear from the StarCraft community: what would you like to see from this years competition, as far as organization, results, and rules? One thing to possibly add is to give the the opportunity to study its opponents. Each bot should have access to the replays of its opponent before the next match. No manual modifications to the bots should be allowed between matches. So, if a bot wants to take advantage of this info it should have an algorithm to study patterns in its opponents play. It would be interesting to see what different things the teams come up with to take advantage of this. A relatively ease thing to check for is the likelihood of the opponent rushing, which could be used to modified the priority of building early defense. If teams come up with successful ways of analyzing the opponent, it will also stimulate the development of less predictable bots. | ||
|
rasdasd
United States82 Posts
On February 08 2011 00:17 Trias wrote: One thing to possibly add is to give the the opportunity to study its opponents. Each bot should have access to the replays of its opponent before the next match. No manual modifications to the bots should be allowed between matches. So, if a bot wants to take advantage of this info it should have an algorithm to study patterns in its opponents play. It would be interesting to see what different things the teams come up with to take advantage of this. A relatively ease thing to check for is the likelihood of the opponent rushing, which could be used to modified the priority of building early defense. If teams come up with successful ways of analyzing the opponent, it will also stimulate the development of less predictable bots. I believe information could be saved between sets, but were wiped between rounds/matches. I believe Overmind saved some flags indicating if the opposing AI would rush or play economic. But i would not allow access to replays between rounds, as this is information that in a regular tournament would never be allowed. For instance in a tourney, people have set build orders as they have practiced them heavily. If a replay was seen before a set was played, the opponent could just create a build order advantage and the coolness factor has just been diminished. | ||
|
Trias
Netherlands53 Posts
On February 08 2011 11:10 rasdasd wrote: I believe information could be saved between sets, but were wiped between rounds/matches. I believe Overmind saved some flags indicating if the opposing AI would rush or play economic. But i would not allow access to replays between rounds, as this is information that in a regular tournament would never be allowed. For instance in a tourney, people have set build orders as they have practiced them heavily. If a replay was seen before a set was played, the opponent could just create a build order advantage and the coolness factor has just been diminished. But this is information that is usually available for public/televised tournaments. A human player can (and usually does) study the VODs of previous matches played by his/her opponent. Of course a VOD isn't exactly a replay, but still contains a lot of information. | ||
| ||
|
|
uThermal 2v2 Circuit
The PondCast
Online Event
Korean StarCraft League
CranKy Ducklings
BSL20 Non-Korean Champi…
Mihu vs QiaoGege
Zhanhun vs Dewalt
Fengzi vs TBD
Online Event
Sparkling Tuna Cup
BSL20 Non-Korean Champi…
Bonyth vs TBD
OSC
|
|
|
EPT
