There have already been a bunch of Youtube videos of BWAPI in action. More are being released every week, and with a year until the actual competition, more are surely on the way. This thread's purpose is to keep all those videos in one place, so that it's easy for someone who's fascinated by this kind of thing to browse through them one after another. If you post new videos in this thread, I'll try to keep the OP updated.
On December 07 2009 02:07 meeple wrote: Wow, nice compilation of all these things. The bot vs bot was interesting, but the purple was obviously much better than orange, no contest.
I don't think bio in TvT is a good idea though. A good mech bot would just wait on his ramp for siege and then slowly push.
On December 07 2009 02:07 meeple wrote: Wow, nice compilation of all these things. The bot vs bot was interesting, but the purple was obviously much better than orange, no contest.
I don't think bio in TvT is a good idea though. A good mech bot would just wait on his ramp for siege and then slowly push.
A good mech bot is also not going to be carrying 300 gas or 1000 minerals
Actually, bio TvT might be a lot better with an AI playing. With some hax micro AI, a ramp break before siege might work as micro benefits the side with medics far more than the side without, thanks to cycling marines (with stim speed specially) without losing them.
On December 07 2009 02:43 SWPIGWANG wrote: Actually, bio TvT might be a lot better with an AI playing. With some hax micro AI, a ramp break before siege might work as micro benefits the side with medics far more than the side without, thanks to cycling marines (with stim speed specially) without losing them.
Problem is that marines, contrarily to mutas can't fly so a ramp break leads to huge pathfiding issues. Anyway a vult only needs two shots to kill a rine ( same for a tank ) so even with a ridiculous micro and medics they are still very vulnerable.
On December 07 2009 02:43 SWPIGWANG wrote: Actually, bio TvT might be a lot better with an AI playing. With some hax micro AI, a ramp break before siege might work as micro benefits the side with medics far more than the side without, thanks to cycling marines (with stim speed specially) without losing them.
Problem is that marines, contrarily to mutas can't fly so a ramp break leads to huge pathfiding issues. Anyway a vult only needs two shots to kill a rine ( same for a tank ) so even with a ridiculous micro and medics they are still very vulnerable.
I can also imagine some sexy SCV repair facility tailing a push making it that much stronger.
On December 07 2009 02:43 SWPIGWANG wrote: Actually, bio TvT might be a lot better with an AI playing. With some hax micro AI, a ramp break before siege might work as micro benefits the side with medics far more than the side without, thanks to cycling marines (with stim speed specially) without losing them.
Problem is that marines, contrarily to mutas can't fly so a ramp break leads to huge pathfiding issues. Anyway a vult only needs two shots to kill a rine ( same for a tank ) so even with a ridiculous micro and medics they are still very vulnerable.
3 shots for tanks. They only deal 50% damage to marines.
Muta vs Marine "micro": (captured with "resolutionhack", unfortunately the video is running very slowly) + Show Spoiler +
ZvZ - Micro module running with a rudimental macro AI vs inbuilt AI (*audioswap because of loud sound) + Show Spoiler +
http://www.youtube.com/watch?v=jSeiL8BxNHw
General Infos: No special Target picking - only closest one (for now) Attack methode - Hold Position -> attack multiple targets in range alse note the new killer build: 7 Ovi - 8 drone scout into 12 pool :-) needs some fixing..
Why use strategic decisions when you can brute force it? Only a fool would throw away natural advantages to use a inferior one. Its like asking Best to go mass drop harass....
Anyway, the muta vs marine micro could still use some work. The most important one would be to find the angle to attack with the lowest density of marines as opposed to running randomly into a blob. It is also important to split the muta into 5 muta groups to one shot marines as opposed to 12 muta groups. The other thing would be to exploiting the bounce angle of the mutalisks to make sure they bounce to closer ones.
As it stands now, I don't think the performance is any better than a higher tier progamer and would utterly get owned by properly blobbed marines that are packed more densely than human micro allows, double plus bad if the marines target the lowest hp muta within range.
The muta vs marine micro is probably one of the least impressive ones, because there are no medics included with the marines! I'm not sure how you could make the mutas micro the marines, but obviously they need to use terrain to their advantage
the tank micro was just absolutely amazing though. No other way to describe it ;_;
Some of these videos are just crazy. While the Wraith and Muta micro videos are pretty cool to look at, I think the Dragoon one is the most impressive (and useful). Coming out of a 12 vs 12 Dragoon battle with 9 goons still alive is hax! Not to mention seeing them all move into position simultaneously is pretty sweet too.
The "whole game" videos suggests the AI designs aren't robust and can get confused at times...... I wonder what kind of decision tree structure is being used...
Those new videos are sweet. It's pretty clear that the AI needs some major work though. For instance, the mine placement is horrendous. I kept expecting catastrophic tank explosions. SCVs seem to be pulled at the oddest times as well. Most of the attacks could have been held off without SCVs helping, and other times, they just kinda wander into the line of fire and do nothing.
I found the reliance on mnm kinda strange. Has mech just not been implemented yet, or does the AI really perform better with mnm?
Looks pretty awesome. Hope to be able to play vs a really good ai one day. Too bad I dont have any programming skills, otherwise I would make a basic hunters zerg and protoss ai. Mindless buildorder plays (normal bw ai) goes pretty far when it comes to practise there.
Right now the EISBot uses a fixed strategy, where certain tech buildings are built at specific supply timings. Also, the placing of mines if completely reactive, based on detecting an enemy. So once an enemy is in the range of a vulture, the vulture will plant mines and then flee. Tanks have a similar behavior and will siege as soon as an enemy is in range and unsiege once an enemy is no longer in range. So this explains why the bot is still predictable.
My long term plans are to have the bot learn from mining replays of professional players. The bot will select a standard build given the map and match up and then adapt to the opponent based on actions that professional players have taken. Right now this is completely theoretical (I'm a PhD student), but I expect to advance the field of game AI.
On December 13 2009 01:33 djsherman wrote: Right now the EISBot uses a fixed strategy, where certain tech buildings are built at specific supply timings. Also, the placing of mines if completely reactive, based on detecting an enemy. So once an enemy is in the range of a vulture, the vulture will plant mines and then flee. Tanks have a similar behavior and will siege as soon as an enemy is in range and unsiege once an enemy is no longer in range. So this explains why the bot is still predictable.
My long term plans are to have the bot learn from mining replays of professional players. The bot will select a standard build given the map and match up and then adapt to the opponent based on actions that professional players have taken. Right now this is completely theoretical (I'm a PhD student), but I expect to advance the field of game AI.
That explains why the bot fell apart to Mutalisks. If the C player didn't know it was a bot I would guess he was very suprised that a player that defended that well to his attack couldn't deal with Mutalisks.
On December 13 2009 01:33 djsherman wrote: My long term plans are to have the bot learn from mining replays of professional players. The bot will select a standard build given the map and match up and then adapt to the opponent based on actions that professional players have taken. Right now this is completely theoretical (I'm a PhD student), but I expect to advance the field of game AI.
I wish you luck. Developing an AI that plays like a skilled human is a much harder (and more interesting?) problem than simply producing a competitive AI.
My long term plans are to have the bot learn from mining replays of professional players. The bot will select a standard build given the map and match up and then adapt to the opponent based on actions that professional players have taken. Right now this is completely theoretical (I'm a PhD student), but I expect to advance the field of game AI.
I'm not sure this is a good plan, since the capacities of a human and a computer is very different. Look at successful chess programs for example, the way a computer plays is very different from that of a human, with different kind of strength and weaknesses. I don't think trying to emulate a human would result in a strong opponent or really advance the field of research that is AI. (since it is probably end up as some decision tree sort type of thing with a lot handcrafting given resource and time constrains)
I think the best way to build a strong game AI in starcraft is to first divide the game up into sub problems (eg. muta vs marines), attempt to figure out which ones can be solved and design AI build orders in things that is good in while avoiding those that it can not cope with.
Build orders are often the major weakness of human players. However for a computer, even something as basic as getting units unjammed or coping with a lone guardian shooting at your CC will take work, let alone things like avoiding a flank. AI is very very stupid and the simplest things to a human is hard in an AI.
Yeah, the D- players were really bad. The Terran in the first video had a terrible build order and then got ran over. In the second D- video, the Protoss put no pressure on the bot and then got ran over. Based on the replays, I would say the bot is currently around 800 (D-) on ICCup. Not quite D level, but capable of beating D- players.
I feel like you could start adding multiple build orders for different AI's, and then make it so that upon scouting certain tech choices the build would change in order to accommodate something like lurkers in PvZ where you need to get obs out in order to deal with them...
However when you add in the micro part of the AI I believe it becomes unfair to deal with as a human so some sort of micro APM cap could be placed on them and then they would become a sufficient training tool to get better against.
Anyways on the comment that the Bot is about 800 or so, I'd say that's a fairly accurate statement, I mean sure we would find D- players that could beat it, but no D player 1500 or so is ever going to get beaten by this thing. I mean the D- toss in the second video has like pure zealot no tech against M+V+T which just destroys all those zeals...
Plus anyone that knows they are playing against a bot can beat it... I would just tech straight to DT's after pressuring with goons... also the AI freaks out when your choke is blocked and they can't see up it.
On December 15 2009 01:13 Traveler wrote: I feel like you could start adding multiple build orders for different AI's, and then make it so that upon scouting certain tech choices the build would change in order to accommodate something like lurkers in PvZ where you need to get obs out in order to deal with them...
However when you add in the micro part of the AI I believe it becomes unfair to deal with as a human so some sort of micro APM cap could be placed on them and then they would become a sufficient training tool to get better against.
Anyways on the comment that the Bot is about 800 or so, I'd say that's a fairly accurate statement, I mean sure we would find D- players that could beat it, but no D player 1500 or so is ever going to get beaten by this thing. I mean the D- toss in the second video has like pure zealot no tech against M+V+T which just destroys all those zeals...
Plus anyone that knows they are playing against a bot can beat it... I would just tech straight to DT's after pressuring with goons... also the AI freaks out when your choke is blocked and they can't see up it.
An APM cap would inevitably lead to very weak AIs. Current AIs aren't extremely good at strategy, especially with limited information and a very rigid time constraint. Having to decide how to use a limited APM effectively would be another hard task to accomplish. Anyways, the whole point of those AIs is to create one that is better than humans and/or other AIs. It wouldn't make sense to intentionally make it weaker. You can always do that afterwards if you really need to.
On December 13 2009 18:30 SWPIGWANG wrote: I think the best way to build a strong game AI in starcraft is to first divide the game up into sub problems (eg. muta vs marines), attempt to figure out which ones can be solved and design AI build orders in things that is good in while avoiding those that it can not cope with.
That results in some very rigid behaviour, though. And for larger problems it may become extremely difficult to map out all the possibilities. I disagree with the idea that bots should be programmed with build orders - it would be far more useful to find generic methods that allow the bot to adapt in a wide range of situations.
On a different topic, djsherman, do you think you would recognize flaws faster if you told the player they were playing against a bot?
I think that T bot in that vs C- zerg video plays very odd. It can't even macro off one base it seems. It just has to not try doing a good strategy and then exploit superhuman mechanics.
Ive played against another BWAPI bot and that bot just kept making depots and units and just runs your over with perfect macro.
I guess you are just less far than that bot. The bot I played could get good building placement on any map. And if an SCV would get stuck it would realize that and use a new SCV. It could also somehow break down any map in pathways.
On January 12 2010 16:13 Glaucus wrote: Ive played against another BWAPI bot and that bot just kept making depots and units and just runs your over with perfect macro.
Had some time (damn SC2) to work on my AI. Now it can handle two standard AIs no problem. The only real challange is the first push, then it's basically cleanup. (This AI does NOT cheat!)
Improvements: -counters earlygame workerharass (from scout-peons) -counters bunkerrushes quite well -kills scout-peons -positions overlords in a more strategic way (still very stupid) -much better targetselections -heavily damaged units retreat from attack -expands (multiple times) as a follow up
APM: 28600
Next testings will be on real players, but until then there is still much to do - Stay tuned (subscribe) :-)
Next goals: -better dealing with high range weapons (goliaths/turrets) and bunkers -refine earlygame BO (mutas could be out 30 seconds earlier) -better scouting -> sunken only when absolutely necessary -maybe implement upgrades in BO -"contain" mode if someone turtles too much -transition into lurker or sth. -add zergling-micro-method
Had some time (damn SC2) to work on my AI. Now it can handle two standard AIs no problem. The only real challange is the first push, then it's basically cleanup. (This AI does NOT cheat!)
Improvements: -counters earlygame workerharass (from scout-peons) -counters bunkerrushes quite well -kills scout-peons -positions overlords in a more strategic way (still very stupid) -much better targetselections -heavily damaged units retreat from attack -expands (multiple times) as a follow up
APM: 28600
Next testings will be on real players, but until then there is still much to do - Stay tuned (subscribe) :-)
Next goals: -better dealing with high range weapons (goliaths/turrets) and bunkers -refine earlygame BO (mutas could be out 30 seconds earlier) -better scouting -> sunken only when absolutely necessary -maybe implement upgrades in BO -"contain" mode if someone turtles too much -transition into lurker or sth. -add zergling-micro-method
Had some time (damn SC2) to work on my AI. Now it can handle two standard AIs no problem. The only real challange is the first push, then it's basically cleanup. (This AI does NOT cheat!)
Improvements: -counters earlygame workerharass (from scout-peons) -counters bunkerrushes quite well -kills scout-peons -positions overlords in a more strategic way (still very stupid) -much better targetselections -heavily damaged units retreat from attack -expands (multiple times) as a follow up
APM: 28600
Next testings will be on real players, but until then there is still much to do - Stay tuned (subscribe) :-)
Next goals: -better dealing with high range weapons (goliaths/turrets) and bunkers -refine earlygame BO (mutas could be out 30 seconds earlier) -better scouting -> sunken only when absolutely necessary -maybe implement upgrades in BO -"contain" mode if someone turtles too much -transition into lurker or sth. -add zergling-micro-method
That is absolutely sick. I'm looking forward to tests on real players!
The botalisk is looking very nice, the target priority is extremely improved from origional versions, Still lots of room for improvement as you mentioned, and I'd love to see it when it's 'done'!
It would also be interesting to see how it fares vs players of each race on B-net, what kind of rank it could get on iccup would be interesting too.
EPT
