The Ur-Quan Masters Discussion Forum

This great game had reborn!!! I've played starcontrol a lot of time. This is the best space universe ever made. I even bought notebook to play starcontrol any time i want, and thats it. I found your project. Many thanks.

It's truly "pimptastic".  Now if someone will make a true SC3...

Maybe timewarp???

When it ever comes out......      :(

Tw sucks. Period.

On second thought..... It won't be so bad, and even good if:
1. The AI won't be as smart as a retard child.
2. There will be a fine plot story.
3. Few/some/lots of/most of the new ships: delete.

But it is just me...

Jiffa, there's going to be a limited amount of ships in the TW Adventure. The Melee ships are sticking, unless you delete them yourself from your .ini folder :P

I'm certain (at least, I hope) that there will be more than "WussieBot" and "MoronBot" to contend with. The ArchonBot is currently in the works apparently, but who knows how good it will be.

As good as ultra-weak cybor in uqm...

Hey, if they could find a way to take UQM's AI and transpose it onto TW, we -might- have a challenge... :P

I had an idea once (while I was in an AI class) to use a neural net in TW, but the amount of variables in melee is enormous.  To get real-time performance, you'd probably need a dual-2.4GHz machine. :)

I also thought about splitting up each button (left, right, thrust, special, fire) into its own NN, and having all 5 produce only 1 boolean value.  That reduces the complexity of each NN overall, but you could only really have 2 NN's run simultaniously even on a machine with 2 cpus, and I'm not good enough with NN's to know what data I could leave out, so I'd have to feed every variable into all 5 NN's.  :P

It would have been fun to set up a server that people log into and just train the AI by playing against it (two AI's would be training at once, one using every command the human player used, the other using the commands that the computer player used.

Archmage is a pretty good coder though, so his archbot will turn out okay.

The only true SC3 would be made by Fred Ford and Paul Reiche III.

And what comes to TW... well you can read my mind. Really.

True they were geniouses, but I think its possible for a game made by other people to be a worthy sequel.

Shades of Leibnitz! It seems like a horrible idea to me. You would have the ship's decision to fire not be based at all on whether it would be using the special, and so on.

I know nothing about programing ai, and am therefor out of my league but I do know that a good ai will need to consider batt when deciding what to do.

That's why it's a hack; but it'd still be better than wussie-bot. You don't really need an intelligent system, only one that acts intelligently (like an anthill).  Initial learning for each AI would take far longer (especially getting them to act coopertiively), but they don't have to continue learning during gameplay (It might be worth the challenge for a Uber-cyborg to learn your moves, but its not necessary). The worst condition I can think of (besides the initial spasms of random action) is if  left and right buttons have the same learning weights; both left and right could determine that it is best to be 1, at which case the ship would move straight forward.  If there was any failure during that learning cycle, then both might think that being 0 would be best (they'd learn the same thing), so the ship still doesn't turn.  A solution for this might be rotating several learning weights.  

This is what random numbers are for.

I'd be interesting to have computer-controlled ships that learned based on tactics used against it. Though coding together such a monstrosity would be (as you have said) quite complex, it'd be worth it. However, I must stress that a computer should only really learn anything new if the tactic is useful. And it should only use one of its learned tactics if none of its immediate alternatives are beneficial.

I have no experience with AI, so I've no idea how one would be able to determine how "useful" a manuvuer really is.

It would probably be much easier if it only took a log during battle itself, and then devised a new strategy while reviewing later. This would learn much more slowly though. Hmm.

That's a really good idea;  you could have a central repository that people send logs of their human vs human games then use that repository to jumpstart the training.  You'd need a _lot_ of logs though.

It is just like a chess engine... How is it called? Yes, Fritz! It learns from its own games and improve. Every move has a "value", negetive or positive(positive = good move). Anyway I have no clue how it works but if someone can look in the code and understand how it works and ofcourse do it in TW engines, it would be great :)

If not, maybe it will give you a direction how to make the engine "learn".

Another idea would be to have humans 'rate' how the computer was doing, manually. Like, "That was a good idea, it was heading for the planet for a grav-whip" or something like that where a failed attempt would be neutral or worse so no learning would otherwise occur.

Again, this would take considerably more time... but it might be fun.