Any autohost for Linux?

[lanctot]

Is there a list of autohosts somewhere? In particular, is there one that works in Linux?

[Licho]

Tehre is one perl one in development, and I plan to prepare springie (with gui-less mode) for linux too on next spring release.

[lanctot]

Do you know who is developping the Perl one? I might be available to help with that or help you with Linux support for Springie. I can't give any promises yet though.. I'm a student and I still need to talk to my supervisor about using spring as the environment to do all my research in. If it passes, though, I'd probably be willing to do most of the Linux support work.

[eriatarka]

What kind of research? Sounds interesting.

[Auswaschbar]

Do you know who is developping the Perl one?

His forum account is bibim, we talked about it here:
http://spring.clan-sy.com/phpbb/viewtop ... 12&t=13778

[lanctot]

What kind of research? Sounds interesting.

Mostly AI research. We've done some work on decision-making by assessing the current game situation via look-ahead simulations. More details are in a previous paper http://www.cs.ualberta.ca/~lanctot/file ... /rtsmc.pdf and a follow-up article soon to show up in AI Game Programming Wisdom 4.

I want to improve on these existing methods by designing an evaluation function for RTS games and I hope to do that using machine (supervised) learning methods on collected game traces/logs/replays as well as reinforcement learning via self-play. With an eval. function simulations can both stop early (rather than simulating until the end of the game) and branch in the middle of their execution, so the estimated value returned by the look-ahead planner would be far more accurate. The result we hope is an awesome AI which can run in real-time by doing short look-ahead simulations to determine what is best to do in the situation it's in.

[imbaczek]

Sounds mighty cool! Out of curiosity - has anybody tried this before with any success?

[Dragon45]

How come you aren't just you using ORTS? You guys wrote it after all...

http://www.cs.ualberta.ca/~mburo/orts/ -

[Kloot]

Indeed, very cool. Spronck's group has done / is doing AI research
with Spring too, but as far as I'm aware they aren't into supervised
learning.

Also, nice to see (someone near) Buro step outside the bounds of
his own ORTS project for a change.

[Licho]

You can help with springie at any moment :) Sources are in spring svn.

[eriatarka]

Mostly AI research.

You know, I was hoping you'd say that

Very cool stuff. I started reading that paper, but got sidetracked, you know how it goes... will try to finish it later. Are you aiming for a PhD thesis with this, or what's your goal?

Man, I wish there was an institute doing that sort of stuff (i.e. game-related AI) at my uni, I'd go there in a heartbeat...

[lanctot]

Ok, some answers. I went skiing for the weekend and was not around computers much.

@imbaczek: Not that I know of, but I'd love to see it even if it's not me that does it.

@Dragon45: Michael Buro is my supervisor, so I'm certainly acquainted with ORTS, but thanks for the link :) The only problem with ORTS with regards to my proposed project is that nobody is playing it every day like Spring, so data collection would be a lot harder. Without a lot of data, there is not enough to build an informative/useful evaluation function. As well, the idea would be to re-apply everything we've learned back into an AI for Spring so that we could measure its performance objectively.

@Kloot: Right, they're doing a "form" of reinforcement learning known as dynamic scripting (and other stuff). Aside from the supervised learning approach, we want to take a more direct approach and do something more along the lines of what was done in TD-Gammon, eg. an agent playing against itself and learning how to act.

@Licho: I'd love to help you with Linux support for Springie, but .. see below.

@eriatarka: Yes, I'm working on a PhD. After speaking with my supervisor, though, the focus will be a bit different than what I was originally aiming for, for very good reasons. BTW, you could always apply to UofA for your master's or PhD. We'd be glad to have more people working on this. :)

So, unfortunately, I don't think I will be able to use Springto do any supervised learning, well at least not yet. The major reason for this is that the proposal is a bit too ambitious. Any game interesting and cool enough for people to want to play it has to be fairly complex with lots of different unit types, an elaborate tech tree, etc. This is a hard domain to get started in because it's too large. I'd be shooting myself in the foot if I didn't start small and show that my ideas work in a small domain first. That being said, we prefer the RL (reinforcement learning) / TD-Gammon approach because it doesn't require human data. Human data is not necessarily "exploratory" in the sense that players are usually biased into their particular style of play which they don't change all that much. RL will explore different strategies (eg. styles of play) and assess which one is optimal. It's much more complicated than this, actually, human data could be better for this very reason... exploring all the different possibilities might be a huge waste of time. But after some discussion we think it's best to stick with this and start small, then work our way up to larger-scale and more complex RTS domains.

My short-term goal at the moment is to work on a automated planning system for Game 3 of the 2008 RTS Game AI Competition happening in August [1]. That being said, I probably will not have much time to work on Spring as I will be spending all my time on that. I will probably lurk here anyway ... and I do plan to do the human study when I'm ready for it. I do hope to eventually put the results of my research into an RTS game that is being played, and Spring is currently the best contender.

[1] http://www.cs.ualberta.ca/~mburo/

[eriatarka]

@eriatarka: Yes, I'm working on a PhD. After speaking with my supervisor, though, the focus will be a bit different than what I was originally aiming for, for very good reasons. BTW, you could always apply to UofA for your master's or PhD. We'd be glad to have more people working on this. :)

Well, that's in Canada, and I'm in Europe, so I don't know how realistic that is I'm currently working towards my master anyway, though my field is Applied Mathematics. I have been considering going abroad for a while after that...

So how's the UofA?

[zwzsg]

Any game interesting and cool enough for people to want to play it has to be fairly complex with lots of different unit types, an elaborate tech tree, etc. This is a hard domain to get started in because it's too large. I'd be shooting myself in the foot if I didn't start small and show that my ideas work in a small domain first.

Have a Kernel Panic evilless autohost! It's a simplified RTS! 6 units, no ressources, etc...

[imbaczek]

lancot: wrt project scale: yes, it's an issue, but only if you're trying to make an AI that plays BA or such. You could take a stripped-down Kernel Panic. It's a simple but still fun game.

Of course, then there's the player base issue - but learning by playing against itself may actually be a reasonable possibility.

PS hah zwzsg beat me to it. well done :>

[lanctot]

eriatarka: UofA is amazing. For anything research related to game AI it's the best place in the world as far as I am concerned. There's a large games research group here, lots of world-class AI and Machine Learning people here. Edmonton's winter is far from appealing, but once you get used to it it's tolerable.

imbaczek + zwzsg: I assume "Kernel Panic" is a mod or something? Question is are people actually playing this game? Suppose we were to setup a Linux-based autohost, when one was ready, and add an option to capture a lot of human-played log traces. Would we get a lot of data? And, what do you mean by "there's the player base issue"? I will look into this more for self-play and see what my supervisor thinks.

[imbaczek]

KP is a mod, unfortunately UF is down right now, but you can still get it through the installer. (I hope.) Player base issue is exactly what you ask - people still think that spring = free TA, which isn't even remotely true in the engine department with all those Lua mods (see Fibre, CA) - so it'll be harder to get a game going, but not impossible.

[zwzsg]

imbaczek + zwzsg: I assume "Kernel Panic" is a mod or something?

Depends on the terminology you use. Some prefer to say Kernel Panic is a "game content" for the engine Spring.

Question is are people actually playing this game? Suppose we were to setup a Linux-based autohost, when one was ready, and add an option to capture a lot of human-played log traces. Would we get a lot of data?

Well, even discarding <100kb demos, I have over a hundred Kernel Panic demos, just for the version 76b1 of Spring. So you'd get some data yes. But how much would you consider alot? Some people used to have KP autohost, maybe if they see that topic they can answer.

I mentionned Kernel Panic because the gameplay is simple, short and to the point, so it's probably easier to get meaningful data out of it than with 200 unit types, hour long, complex games.

[lanctot]

What is "UF" ? Does Kernel Panic use any of the original content from TA which makes it illegal to play if you don't own the content.. or is it all free?

By "demos" do you mean videos of games played? I would need replay files in a file format that can be automatically parsed. When I mean lots of data, I mean *at least* a few thousand games, preferably hundreds of thousands, and ideally even in the millions. Everything would have to be automated. The data collection would be done via the autohost automatically after every game: replay files would be sent to the host and stored. I would write something to parse these to extract the information I need. This is why I asked Tobi to give me an estimate of how many games are played per day, but he hasn't replied yet. Does anybody know the answer to this, even roughly?

[lanctot]

Well, even discarding <100kb demos, I have over a hundred Kernel Panic demos, just for the version 76b1 of Spring. So you'd get some data yes. But how much would you consider alot? Some people used to have KP autohost, maybe if they see that topic they can answer.

I should mention that 100 games would be good to start testing some of my code to pull out stuff I need, but I would have to take a closer look at KP and talk to my supervisor about it first.