Neural Network Trained Enemies

http://cg.skeelogy.com/training-neural-networks/
The enemy robots in this mini-game uses neural networks (NN) to do decision making. Prior to what you see here, the NN has undergone some training so that it has learnt how to react based on a human player's reactions.

Notice these things about the trained robots in the video:
i) They approach me when I'm near and stop to attack when they are in shooting distance. They approach me again when I back away.
ii) Once their HP is low, they know that they should flee instead of trying to attack again
iii) Although they have started to flee, they try to attack me again when I'm not facing them. Once I point the cursor at them, they know that they should flee, even though I haven't even started shooting. This implies that they are especially good at back attacks.

And all these are done by simply training the neural networks.

Note: The fleeing in a zig-zag manner is procedural and not done by the neural networks.

Category:

Entertainment

Tags:

• Neural
• Networks

[ahmadmagad]

OK !, i loved two parts, first how the AI enemies change their current state from flee to attack once U R in the range n nt concentrating on them, 2nd how the robots smartly flee away depending on your firing direction n for sure their health state, i didnt see your designing Diagram but i just make a guess, n i didnt notive any more features plz post back if i missed somthing n i used to work using Fuzzy logic n i m begineer in NN, so is true u can make them learn !

[skeelogy]

Yep it's true that you can use NN to make the A.I. learn! Since you are working on NN as well, you might want to take a look at my other project called A.I.Fighter. It's a game where the A.I. evolves based on the way the player fights.

[essaied66]

nice work, I just have a question, how can you test a trained network and what format do i need to save the trained network in?

[skeelogy]

One way to test the trained network is to just let the AI use the network and see if the robots behave as expected. A more quantitative way will be to obtain more samples (not the ones that you used to train), with both inputs and expected results. Then feed these samples into the trained network and see if it gives you the same output as the expected results.

Since the network is just a series of values (weights, no. of edges, no. of nodes etc), it can be saved simply as a text file.

[GBart]

Cool. Now make them compete with each other over resources and reproduce with very slight variations.

[egokick]

I hope skynet works this way :P

pew pew *FLEE*

[mutehero7]

I personally think learning AI is the future of gaming. The accessibility of online gaming makes for plenty of opportunity to learn... Can you imagine learning COMs in Halo, Smash Bros, or Burnout? I'd be scared.

[egokick]

@dubdubhate no it doesn't. A neural network is just a mathematical function that attempts to emulate how a neuron works, so they can learn things by trail and error. They are pretty simple at the moment in comparison to the human brain which has trillions of neurons.

[rkyeun]

@RoyalSeducer87

Gödel's Incompleteness Theorem isn't a barrier to human-like A.I. because we're subject to it too.

[NicholasCobalt]

@BenMcCormack91 Allowing AI to develope in this way, then hard coding the developed experiences into a bot without limiting future experiences would work, but enough experience with a dumb player could make the AI less intelligent, alternatively enough experience with the best players would make the bot impossibly difficult.

[rerere284]

I think they need to add in them socializing with eachother so they coud be more like humans.

like wanting to protect their pals too.

unless there zombie nn.

then they would be like these.

[macca1088]

I would love to see AI in a game that learns from the players actions and adapts its play style to meet them...I would love to see this type of AI in FPS and RTS gaming where not every enemy soldier is cannon fodder, throwing themselves at you - they have a sense of self preservation and will run away...

AI like that can be seen in S.T.A.L.K.E.R's A-life, although I don't think that uses neural networks - more hardcoded so that if the panic threshold is breached - run away!

[BenMcCormack91]

Well, for it to be as smart as a human, theoretically, it would have to learn experientially from birth, as humans do. In other words, to get a smart AI, it would have to go through life from the beginning, picking up experience as it goes. Hard-wiring basic functions into it as a basis is probably the best you could do for practical gaming, without creating an entire past life in an artificial world for the character to understand.

[RoyalSeducer87]

Nope, and probably we will never master it. Look for Gödel's incompleteness theorem and Lucas' argument...

[dubdubhate]

Doesn't Neural Network means a A.I. that is as smart and cunning as a human. Have they really master NN yet?