|
On October 02 2011 08:05 firehand101 wrote:
You have done it. You have successfully found a place where ladder fearing noobs can retreat and practice just as well, without losing any ladder points! i thank you for this
:D. I realize you're joking...but joking aside, playing ladder leads to dealing with really unpredictable strategies (at lead it does in lowly diamond). For me, practice is most effective when I can try the same thing over and over again. GT AI allows me to do this so I can then ladder and deal with the unpredictable while having my chops down.
Thanks!
|
Hi
Thanks for the map, very well done. I use the method 1 - download and play from EU custom game I have a question -
1. Is this updated for version 1.4 - Xelnaga/Shattered Temple / Metal?
2. Is this only vs Terran I can play?
3. I want the AI to learn a BO that I face vs terran - marine opening into BF harass into marine tank timing - how do I get the AI to do this?
|
On October 02 2011 07:59 Destructicon wrote:
I like the concept and principle of this AI, however I have a problem with it, one small fault that keeps me from getting it.
I don't like the fact that your AI gets map vision and extra resources at higher difficulty levels it kind of detracts from the already impressive accuracy it has to mimicking real players because of the diversity of strategy it uses.
It would be nicer if at higher levels of difficulty the AI uses either harder to execute strategies and/or executes them better while also increasing the APM it does and also learns finer micro, like constantly kitting zealots with marine/marauder till the zealots are dead then kill stalkers, or if it does efficient drops or poking with vikings etc.
I would love if you could develop and AI that works like that, doesn't cheat at higher levels but just plays better overall, would feel like Insane is GM/Master, Very hard is Master/Diamond, hard is Diamond/Plat.
Great idea. I find it obnoxious that the AI sees all my shenanigans from the get-go and hard counters the crap out of me in Hard~Insane on the mid to late-game scenario, mainly because it didn't even have to do any scouting at all, while I struggle all around just to get info. There should be an option in the options tab for you to enable or disable the AI's map hack.
The Zerg AI also uses too much APM maneuvering each individual overlord to safety when an attack comes to its' main that uses up so much memory due to it going at 1000 APM, which is not only physically impossible for a human player to do, but is also impractical. Toning down the AI's max APM via the options tab also suddenly makes the AI dumber for some reason, often times with the AI not being able to do its' own build orders because it needs at least 600 APM minimum before their actions register. Not to mention, blocking an expand via a probe or SCV results in the AI one or two-basing for a good minute.
All in all, the AI still kicks my butt, but sometimes only due to the map being fully revealed to it, instead of a really good army composition or constant effective harasses.
|
On October 02 2011 07:59 Destructicon wrote:
I like the concept and principle of this AI, however I have a problem with it, one small fault that keeps me from getting it.
I don't like the fact that your AI gets map vision and extra resources at higher difficulty levels it kind of detracts from the already impressive accuracy it has to mimicking real players because of the diversity of strategy it uses.
It would be nicer if at higher levels of difficulty the AI uses either harder to execute strategies and/or executes them better while also increasing the APM it does and also learns finer micro, like constantly kitting zealots with marine/marauder till the zealots are dead then kill stalkers, or if it does efficient drops or poking with vikings etc.
I would love if you could develop and AI that works like that, doesn't cheat at higher levels but just plays better overall, would feel like Insane is GM/Master, Very hard is Master/Diamond, hard is Diamond/Plat.
A laudable goal.
There's a reason AIs don't "play fair," which is that is very hard to make an AI that can out-think a person. This isn't just true for green tea, but for any computer-based AI. In fact, as I'm sure you know, there's an entire field of Computer Science dedicated to trying.
I'm sure everybody would like to have a fair-playing, GM AI waiting to play them at any time. I also find the "ohh i see your observer, guess I'll just park my observer on it forever" types of gameplay frustrating since they don't mimic reality, but the truth is it's just the way things have to be to make the computer a challenge for us, at least for now.
If you're really that bothered, pick up the torch and consider a career in computer AI research. 
|
I just wanted to say thank you to those that work on this. I lab this thing about 20 games a day and usually still lose a few. My Masters friends think its terrible and I should ladder, while my Gold and lower friends thinking I'm like a pro. This keeps me SORT of sharp (at least mechanically) with none of the stress of laddering. TYVM
All your work is appreciated, and GL making this thing even more of a pain in the ass <3
edit: Since its the topic of the moment- "cheating" and unit countering are fine with me, it's the human equivalent of scouting- since the AI really cant think. You can still out smart the hell out of it. I spent a LOT of time making AI's with WC3... which had a great interface for it. Once you've spent hours with them, you realize there are serious limitations to what they can and can't do.
|
4713 Posts
While I do understand that something like, unit preservation and effective harass, is something very hard to code, I didn't think it was that hard to make an AI just incrementally get more APM/micro per difficulty and also broaden the range of strategies available to it.
I'm not asking much, just for the AI to not cheat.
|
I disagree with the people saying there should be a no-vision AI at hardest difficulties.
Starcraft is a lot about predicting your opponent, getting a solid idea of what they're getting without seeing much, and even just guessing and/or unintentional lucky events.
When practicing and/or facing an AI that is (at least currently) weaker than masters level players, or even diamond, it needs all the luck it can get. Having the AI know what you have and know where you are teach players that they need to be prepared for their opponent (prepared for the worst-case scenarios), and not rely on lucky counter-attacks or lucky harassment.
However, I think that currently the AI doesn't even counter what your getting that well at the moment (not sure haven't tested too much and/or don't remember testing I did do). Things like getting bunkers and marauders vs a roach rush, turrets or cannons or observers for Dark Templar or banshees, thors or phoenix for mutalisk, etc.. In fact, that's something I think really needs to be worked on for Green Tea AI that could make it a whole lot better — following a BO, but before even finishing the BO, modifying it a bit (such as detection), or switching to something completely different (like double stargate instead of 2 gate expand) to counter the opposing player's strategy.
Mineral cheating I don't like though, since it's a completely different problem. That said, I could care less if mineral cheating is in ONE difficulty (preferably with player-adjustable income multipliers) such as insane, which is what Blizzard does.
While in an ideal world it would be nice to have an option to be able to disable CPU vision and have it play well, but the problem is that disabling revealed map for the AI results in it being terrible at countering absolutely anything, and scouting is an extremely difficult thing to program effectively into an AI. Humans know generally where the best expansions are to scan, or where they might be hiding tech, or when to gas steal, or all sorts of other things, but without specific programmed guidance, the AI cannot do that. Overall, it makes an unnecessarily easier AI at the expense of a lot more programmer work, which is not worth it until the AI gets good enough to beat most master players.
And if any of you did not know:
To play a non-mineral cheating AI, Hard is the best one — the AI itself doesn't improve past hard.
To play a non-map-cheating AI, Easy is the best one.
|
Until today, I had been playing a GTAI that worked fine on 1.4
The ones uploaded by Humdinger (GTAI 104a Melee, Co-Op, vsAI & Custom) maps were working just fine. In fact, they are FAR more challenging on Medium than the Current ones by IcyTournesol.
The new ones are actually nicer, and I like the Aggressive vs Defensive mode options. Very nice. I had been playing the 104a w/o the New Strength Analysis on (it slowed my computer down WAY too much on like 4 bases each)... but still provides a much better challenge than the current ones recommended.
I don't know why this is... but just wanted to give some feedback while this thread is hot. The only other thing I would say is maybe prioritize the Rocks on maps a little higher, especially on XelNaga and Shattered Temple. The thing seems to trip on the rocks quite a bit. Lastly, maybe assigning Map Specific Zones to the AI to build structures in for the Cannon rush and maybe even like a Proxy Rax. ie. Behind the mineral lines of naturals or in smart zones in the main. I LOVE that the AI does these things, and get a giggle every time.
Again, thank you for your work I <3 this.
|
Is it possible to improve some of the building placements? like for example full walloffs with FFE in PvZ or depot-rax-depot walls in TvZ. of course I preface this with saying I know nothing about coding SC2 custom maps, but i shouldn't be able to run lings past the choke points so easily.
|
It is possible, but it's difficult and takes a lot of work to do. Several months ago I heard of a guy that managed to code an algorithm to make walls at chokes with just depots or something (and I think it still had some problems), but not mixed buildings like barracks, which is more difficult. I haven't heard from him since, I don't know if he's still working on it, but I don't think so.
You should just pretend the computer walls off at x time though, or enable the "kill units when they go up ramp" feature the maps have, which I think is super dumb.
|
Wow! can't realise I havent seen this before! well done! will be the first thing i try when i get off work!
|
|
|
Now that season 4 maps are confirmed, can IcyTournesol please put Metalopolis (w/o close spawns) back into the rotation of GTAI maps that are kept regularly updated?
Thanks so much!
|
The AIs don't even seem to do what they say they do.
A build labeled tank did like a 7 marine +1 tank attack, and then got marauders, marines, medivacs, and some vikings and hellions(along with a few tanks) — how is that build tank? I would consider a tank build to be marine-tank, which I suppose he did, but it was only 1 tank and like 7 marines before getting other stuff.
There was no reason to be getting vikings or marauders or medivacs either, since I just massed mostly zerglings with some banelings. If it switched to automatic play, shouldn't it be able to deal with just 1 or 2 units better than it did?
On July 15 2011 21:47 ptanhkhoa wrote:
_ Also update the new version that change the strategy core. After the first strategy failure, instead switching to random commander, it will switch to general purpose commander that can counter to any strategy. But it need more testing and feedback.
Is that why it did that?
Getting a bit of every unit is a terrible idea until/unless it's late game like 20 minutes deep. It's definitely not the way to attempt to counter any strategy. Is that sort of behavior still around? it seems like it.
I also selected a 14 hatch 14 pool commander, but [the 3rd game] it did a 7 pool instead — is that intentional?
|
United Kingdom1381 Posts
Hi, I've been messing around with the 1.17 version on the EU ladder and its much improved!
The two big bugs that stand out are:
- The AI never breaks the rocks down for expansions and faster pathing between bases, sometimes it expands next to me instead of taking the safer gold.
- Workers seem to glitch out when an enemy unit comes within a certain proximity and get stuck. E.g. leaving a marine at a watchtower caused 30 drones to glitch while trying to transfer, or putting a unit behind the mineral lines prevents all mining.
|
|
|
Great map where I've practiced on alot! Too bad there are a few bugged maps on the server now, especially the 'b' or 'c' versions (example Xel'naga GTAI 1.14b) where there are no larva's for zerg or just no option to select cpu's buildorder.
I'll try out the new 1.17 version of the map. Still, great job on this map and it really helped alot! I know a Terran player who was bronze (with a little ladder anxiousness), began to practice alot with the GTAI and when he started laddering again he instantly climbed to Platinum. Keep it up!
|
Tried it for the first time today, very impressed. I tried a new build as Zerg vs Very Hard normal AI toss and just completely owned, I didn't even have to think, scout or react, just macro according to the build. Tried vs medium GTAI, a commander using a ground force, and I actually had to work for it, he attacked way faster than the very hard AI, at a time where my build really didn't have any defences.
To have an AI which can do realistic builds in a decent manner? Now that's nice.
|
I wanted to give a try to the AI's 4 gate in PvP on Tal'D'arim, but I must say I'm a bit perplexed at what I saw. When I scouted him, I saw he took 2 gas and didn't reserve any chronos for warp. How is that a 4 gate ?
|
On October 03 2011 05:11 SCbiff wrote:Show nested quote +On October 02 2011 07:59 Destructicon wrote:
I like the concept and principle of this AI, however I have a problem with it, one small fault that keeps me from getting it.
I don't like the fact that your AI gets map vision and extra resources at higher difficulty levels it kind of detracts from the already impressive accuracy it has to mimicking real players because of the diversity of strategy it uses.
It would be nicer if at higher levels of difficulty the AI uses either harder to execute strategies and/or executes them better while also increasing the APM it does and also learns finer micro, like constantly kitting zealots with marine/marauder till the zealots are dead then kill stalkers, or if it does efficient drops or poking with vikings etc.
I would love if you could develop and AI that works like that, doesn't cheat at higher levels but just plays better overall, would feel like Insane is GM/Master, Very hard is Master/Diamond, hard is Diamond/Plat. A laudable goal. There's a reason AIs don't "play fair," which is that is very hard to make an AI that can out-think a person. This isn't just true for green tea, but for any computer-based AI. In fact, as I'm sure you know, there's an entire field of Computer Science dedicated to trying. I'm sure everybody would like to have a fair-playing, GM AI waiting to play them at any time. I also find the "ohh i see your observer, guess I'll just park my observer on it forever" types of gameplay frustrating since they don't mimic reality, but the truth is it's just the way things have to be to make the computer a challenge for us, at least for now. If you're really that bothered, pick up the torch and consider a career in computer AI research.
One workaround could be to continually update what the AI does whenever there's a shift in metagame.
|
|
|
|
|
|
EPT
