A.I. Fighter

http://cg.skeelogy.com/aifighter/
A.I. Fighter is a 3D beat-'em-up style fighting game in which the player will control a fighter using the keyboard and engage the computer's fighter over a short duration. It is created for us to test out training of computer A.I. using Neural Networks.

Category:

Entertainment

Tags:

• Artificial
• Intelligence
• A.I.
• Fighting
• Games
• Neural
• Networks
• Finite
• State
• Machine
• FSM
• Combos

[Skinnyram]

Did you use an engine that is avaliable to the public? I'm wanting to do Neural networks for my final project but having problems locating some code thats relevant.

[skeelogy]

Nope, it wasn't done using any engine. Some basic source codes can be found in neural network books I believe, so you can try digging into those.

[DrJBN]

Great vid & idea. Backprop is prone to "catastrophic interference" Does your player have trouble initially retaining old skills when learning new ones?

[skeelogy]

Yes it does actually. It'll tend to forget some of the combos but that's kind of expected with the interference from other samples, like what you have mentioned. Seems like you have done quite a bit of research on this topic as well :)

[futurehorizon24]

From the way you described it this game is no Neural Networks but just simple if statements. Dont post things like this when there are people actually trying to develop real methods.

[skeelogy]

It's kinda strange that you wrote that there's no neural networks involved. Have you watched the whole video? We have another version of the game which is running on Finite State Machine that uses if-statements. But that is used as a contrast and is not the feature of this game. Perhaps you can watch the later part of the video which talks about the neural networks.

[red15nls]

Hmm, the time it takes us humans to figure out how to do things takes 10 to 20 years and sometimes even longer. Which we call growing up :) So the AI needs to grow up too. Just like you need to learn a new game the ai has to learn your abilities and how to react to them.

[Althar93]

Cool, just what I needed : an explained implementation of neural networks!!

You can find plenty of stuff on networks, but they never talk about a specific implementation, rendering it impossible to fully understand.

Thanks for the vid, and awesome work!!

[drowjam]

The biggest problem i have noticed with neureal nets is that they can be trained to do on thing but if you train them to do another they will often not be able to complete the orginal task they were assigned. I had an idea for an AI i once that involved using neural networks to search relatioanal databases to reconize patterens in a players ablility but that was just a little to much for me. anyway great vid.

[EZbakeROFLcakes]

You could try making advance AIs in UDK :D

[dajwilkinson]

@Houshalter Maybe every time AI loses, its last move (or set of last few moves) is the basis for the back-propagated error signal? Simple game-learning by this technique has been around for decades - e.g. Martin Gardner, Further Mathematical Diversions (1969), chapter "A Matchbox Game-Learning Machine".

[xistoxs]

Thats awesome. I hope that NPCs become super intelligent really soon.

[rcparts]

@virgildog1950 It could have been done the way you said (like TD-Gammon did), but they've used a different approach. It's on their site.

[virgildog1950]

@Houshalter Keeping your health up and hurting the player reinforces the bot's learning because it complies with its objective.

[TheReasonWhyGuy]

you should have 2 train against each other...

[CyberWalker2004]

Really nice background music!!!

[Houshalter]

But how does it know what action is best in a given situation? In order to train it, you have to have an example of the best possible answer in every situation.

[Ambakapu]

I like your story telling method .)) fun! great job!