I want your replays

[SpliFF]

Hi guys,

I'm working on an AI that learns mods by collecting stats from replays. I have dl'ed a lot of replays from springfiles and Spring Replays but I need a lot more.

Any game played in the last year is fine regardless of map or mod.

To send me replays go to your Spring/demos folder, sort by size and zip up all the smaller replays (some replays get enormous) into a single zip, bz2, rar or 7zip file.

To send replays:

Attach the archive to your post (using Upload Attachment below). I'm not sure this forums max upload limit.

For really big archives use RapidShare (up to 200MB)

I can receive files via instant messaging. I'm spliffim on Hotmail, Yahoo, Jabber.AU, MySpaceIM and IRC Freenode.

email it to spliff/at/warriorhut.org

Your help greatly appreciated.

[smoth]

If I may ask, how does the ai handle version changes? Does it have to start over again? What if there are major changes in the way things work between versions?

[Jasper1984]

This does sound awesome. Before i ask a million questions and make myself look even more stupid, i will ask a few:

• Did you read about some statistical/AI method(s) that you found really cool, and are you overestimating its power now? Hey, that happens to the best of us! What kind of amounts of recordings and cpu time will the AI need? Can you somehow give the AI hints?
• Assuming the first one is no, will you write a paper or something on how it will works?

(non sarcastic)Good luck.

[TradeMark]

why smaller replays? they are mostly the replays which didnt start at all = broken replays.

anyways, i think its pretty pointless to make AI that way since if mod changes, the AI needs to learn it again... and since mods changes pretty often, you need to make it re-learn very often. and mods like CA changes every hour, its 100% pointless to make AI for CA.

[Jasper1984]

I suggest we know our criticism back, i certainly plan to. If the OP wants to, i think he certainly should try. I only stated some negative possibilities so he could consider and take them into account, as overestimating approaches to problems happens.
Remember, we know next to nothing what he is planning to do. For instance, trademark, his approach might need a full re-computation when the mod changes a little.

[smoth]

I didn't criticize I was just asking a question.

[SpliFF]

I just mentioned smaller replays because I have an 800MB one for some reason. Normal games should be up to 3MB.

I forgot to mention I can receive files via instant messaging. I'm spliffim on Hotmail, Yahoo, Jabber.AU, MySpaceIM and irc Freenode.

Feedback is good but I'm not going into this naively. Basically the stats I'm collecting represent the average power of a unit vs. another unit type and also vs. a generic armour type (if the mod has multiple types). I tried to calculate these using a formula (damage_factor * aoe_factor * reload_factor * burst_factor) but balancing the factors is too hard, particularly across so many mods, and it's incomplete anyway since projectile accuracy, turret turn rate and other things also factor in. There are just so many variables that can affect REAL combat performance that the unit properties don't really tell you.

So I came at this from the same direction human players do. Watch enough battles until you get a 'feel' for what works. The calculation is much simpler too because for each unit or unit/armour pair I only need average damage per hit (total damage dealt / number of hits), number of hits, and the number of that type built. This still leaves out accuracy but since the engine doesn't callback when a weapon fires and misses then an approximation based on the number of hits vs. the number of each type built will have to be used. Still, it's a fair bet that if unit X and unit Y are built often but X never damages Y - then X is not an effective unit against Y spam. The best thing about this method is that it is not mod-specific - it should work for practically any mod with fixed unit and armour types (ie, every mod except micro modules)

I split the stats into one collection per mod/version/map combination. I will probably average stats across the last 2 versions of a mod (last 6 for CA) so the AI can deal with new versions, even if this reduces accuracy. I can also fallback on global mod stats when map-specific stats are unavailable.

The main thing to realise is the calculated 'power ratings' do not need to be 100% accurate. They are a guide that hints to a given unit when to fight or run. If a unit type kicked its arse in the past then it's reasonable to assume it might happen again, except in very extreme re-balancing cases. The AI works on the same limits human players do so if it can't see a threat it can't calculate it. But when it sees 2 mixed attack forces it can calculate attack and defense ratings based on the unit types and decide what to do. If it scouts a player building bombers then it can search its memory for the most effective anti-bomber units and build enough of them to counter the threat.

The per map stats are used to tell the AI how popular a unit is for a given map and/or mod. If nobody builds it then there is a good chance it isn't worth building. The map stats also provide the AI with popular unit choices and base/defense locations.

[SpliFF]

re: jaspers questions

I can't tell you CPU requirements until the AI is more complete, however it's intel updates seem to complete very quickly and they only need to run every few seconds.

Yes you can give the AI hints. In fact it runs as a widget so you can actually choose AI control and priorities on a per unit or per region basis (select an area of the map and choose 'defend', 'build base', 'capture' or 'scout'). In other words, this is a helper AI not an opponent AI, yet it is capable of playing almost unassisted.

The whole idea of helper widgets bothers some people who play Spring like it's an action game rather than a strategy game. I don't agree but it's likely the final version of this AI will be part of a new mod, if for no other reason than so I don't have to hear any whining about 'cheating'. It's purpose is not to be a cheat, it's to micromanage idle units that would otherwise get slaughtered because they're too dumb to recognize threats and opportunities. The replay scanner is designed to build a database of potential threats based on previous games. This is exactly how human players get their data too.

[Day]

PM me in lobby PRO_Day

[panzeriv2]

Nice idea

[TradeMark]

I just mentioned smaller replays because I have an 800MB one for some reason. Normal games should be up to 3MB.

800meg replay? wtf. really, my biggest replay is 7megs. i have 1200 replays and they take 670 megs total :D

[AF]

Uber huge replay -> [Zip archiver] -> Teeny weeny baby zip

repeat and rinse

[maackey]

I have no idea what are in these replays, I just picked the ones that were not too small and packed them up.

Good luck on your project!

[Hobo Joe]

http://www.mediafire.com/?sharekey=9748 ... b9a8902bda

Good luck with the project.

[SpliFF]

Thank you to maackey, hobo joe, satarik, quantum, Day and Hoi for your efforts.

Now some bad news.

It's become pretty obvious to me that the AI widget these replays are for is going to upset a lot of people. Even though I totally disagree with their reasons I really can't be bothered with the controversy and bad blood.

It's hard to know exactly what the level of concern would be without actually releasing it, at which point there would be no going back.

At first I just didn't care. There's certainly no legal reason not to do it and I'm pretty good at ignoring whingers. However I really don't want to spoil Spring for anyone and many players seem comfortable with playing the same game over and over (basically BA on DSD). It's these players who are going to scream blue murder when a widget comes along to change the balance of power or gameplay. Just look at the outcry when some minor unit stats change and multiply that by a million.

Now the good news.

Yesterday I decided to start my own mod. Since I'm the one doing the work, and no players currently exist, there is absolutely no way anybody can claim I'm spoiling their fun with AI automation. Those who prefer TA's twitchy gameplay can continue to twitch away on BA or whatever while I build a mod based on complex strategy and AI widgets like this. Naturally players will be able to turn automation off but they'll have absolutely no grounds for expecting others to do the same - as the widget will be a standard feature of the mod.

I hope those who really want this widget for another mod or took the time to send me replays will understand my position and not feel too pissed off.

[Tribulexrenamed]

Thats just an awesome plan. More lua for me to steal and put into my mod.

Ignore the whiners. They will always be there.

[Crayfish]

I approve, incidentally. Three reasons.

1. Not everyone likes micro and there are some occassions where you just can't micro two things.

2. A widget will never be as thoroughly sneaky as a human player could be, so it won't replace human input, it'll just allow you to concentrate on the important stuff while neglecting the 'B tasks' less.

If you're whining about people cheating by automation, isn't that kind of admitting you're less good than AI?

3. If you don't like it you could always turn it off.

[SpliFF]

2. A widget will never be as thoroughly sneaky as a human player could be, so it won't replace human input, it'll just allow you to concentrate on the important stuff while neglecting the 'B tasks' less.

This unfortunately is the key to the whole issue. Most RTS doesn't have any 'important stuff'. TA is basically doing the same 'B tasks' repetitively until you are able to do them faster than your opponents.

'Strategy' is this scenario is basically deciding where to place your defense lines and when to tech up. In fact CA even takes that one out.

It's very hard to be sneaky in TA. Scouts and radar are cheap and even cloaked attack forces are usually spotted before they reach the enemy.

I think it's fair to say that most game-ending strategies are only viable if you already out-produce your opponents. Once the gap is wide enough it doesn't really matter what you spam. Build enough of anything and it'll eventually break through.

The difference with my proposal is you have defined objectives so your choice of units and tactics have a greater effect on the outcome. At this point your statement becomes true since the AI won't be very good at deciding your tactics for you.

[Sleksa]

You think that a AI could do better micro than a human player?


With or without support AI's youre still going to lose to top players, no matter what mod ~~

[Google_Frog]

Strategy is basically what you spend your resources on and where you put your stuff. These are the choices a player makes.

It's not just where you put defence and when you tech up. There are other units to place and many more things to spend resources on.