Robo Home | Changes | Preferences | AllPages

A bot by Christian Schnell, which is controlled by a person using a joystick or the keypad. You can donwload it from http://user.cs.tu-berlin.de/~lulli/hcbot/

This is great! I'd never heard of it before today. I'm usless at it, but that's no suprise, i've never been any good at driving games etc. -- Tango

/Highscores, anyone? ;-) --David Alves

Hmm, has anyone tried hooking up a Neural Net to this thing. Like teach it to play how you play... Sure you have a hell of alot more input than the bot would, but it would be fun to watch. -- Jokester

Seeing as all your decisions in playing are generally 'move just enough to avoid the bullet I can plainly see', unless the net had bullet positions as an input, you'd at best train it some patterns for messing with a specific bot's gun. Of course if you had bullet positions I guess you wouldn't need the neural net. Though to be fair, if you look at something simpler, like saving your pattern against a high-end stat gun and then having a nano just follow that pattern precisely, you could end up with a flat and tiny movement. That's one of those things that I've heard many people say, but nobody's really done any research into afaik. So if you were looking for an HCBot project you might want to look into generally recording how you avoid a stat gun and then making an engine for generating flat movement based on certain moves, or some or all of the recording. I just don't really think anything you could feed a neural net of feasible complexity would be helping a whole lot.

EDIT: Oh, yeah, and now that I think of it applying it directly probably wouldn't help a whole lot. I recall either nano or kawigi doing a test where they graphed way too many rounds to have human patience for as HCBot. Movement was nowhere near flat or useful when generalized directly. If the other bot chooses one different bullet fire power or has a different decay rate in their stat gun, you're done for in a massively bad way, because the baiting won't be timed right. Just to correct myself a bit, but I'm sure certain elements of the movement can be spliced up and generalized if you want to play with it. -- Kuuran

Hmm, well one big obstacle stopping me from trying this is that I am terrible at controlling this thing (I can barely hit sitting duck and get destroyed by the other sample bots ;) ). What could be very cool (although would take alot of patience), would be if you ran some battles at a low frame rate and had the net apply the movement inputs you give it to the information it is getting from regular events. Maybe get the bot to learn how to dodge bullets... Definitely not very applicable, but could be kind of cool. -- Jokester

This is very interesting although I will have to try it at home. I have, for the longest time, wanted to be able to do this. It would also be a pretty good way to teach a robot to learn, besides the obvious, because it provides a way to input data from the user. It is certainly something that might become the next type of robot. -- Kinsen

I just had an idea for a very cool modification for robocode. What if you could see from the tanks perspective (a battlefield view from the radar)? If you could do that then you could create a very interesting game, would the human driver be better than the AI. You could use both a HUD and logical displays to represent the scans, and while the human would have the better mind, it would be at the disadvantage of movement (although it could then be modified for a mouse look for radar/gun and a keyboard for movement). Oh the possibilities -- Jokester

If you could only see from the tank's perspective then it would be extremely hard for the human driver to win. Yes, the human driver and the AI would be equal but then the AI would have the advantage because it has near-infinite memory capacity while the human driver would probably forget all about the tank that was scanned a while ago but was now extremely close to it. -- Kinsen

Sounds a bit like this game: http://www.freearcade.com/Battletank.jav/Battletank.html I can't decide whether or not I'd pick the AI or the human to win. The AI would be very difficult to hit, because I don't know how many of us can improve much over either head-on or linear targetting. The only advantage the human would have is being able to see the bullets. Oddly enough, this caused me to imagine a little man in each of those tanks on the Robocode field. It must really suck not to see the bullets coming at you, all the while knowing that they're coming slow enough to dodge. And then bam - explosion. -- Alcatraz

I made a modified version with a seperate gun control but it doesn't work with a joystick since I don't have the resources to. You can get it at http://www.geocities.com/seakinsen/cs.HCBot_Modified1.00.jar. -- Kinsen

Robo Home | Changes | Preferences | AllPages
Edit text of this page | View other revisions
Last edited June 14, 2005 22:17 EST by Kinsen (diff)