AI Only Skirmish Mode

[Guest MachineDoctor]

Just starting to tinker with the AI, I noticed something that was omitted that really would help test new AI Scripts.

 

It seems that in Skirmish Mode you MUST use a Human Player. You cannot pit 2 (or more) AI players against each other.

 

I really think this option would be of great benefit to everyone who intends to mod the game, create new AI Scripts and even map builders.

 

Maybe when they start patching again, they will allow you to select an AI player as player 1.

 

I welcome any comments or thoughts on this!!!!

[Koen 2]

Hmm...

what if you make a single 'unbeatable' unit for the Human player (it's dozer) and then have it join one of the AI teams?

I'm not sure what happens though when the AI attacks something with 'infinite' HP...

[Guest bws]

Well... when you want to make an unbeatable AI you gotta use neural nets

[Guest MachineDoctor]

You guys are taking it a little farther than I had imagined.

 

I was just looking for a way to simply observe 2 AI players playing against each other.

 

This would be good to check out maps (bottle necks etc.) and how good the AI scripting is.

 

The unbeatable unit joing an AI team might work (I guess???).

 

The neural network is WAY above my head. I hope that you were being sarcastic BWS!!!

 

Thanks for the replies!

[Guest bws]

Neural nets are easy to make.. really...

 

 

This is how they work:

 

 

There are 10000000000`s and 1000000`s of little neurons in a neural net.. evey neuron has got an input and a output and a target.. a neuron has to adjust its homemade formula so that the input becomes the output like in the target... when you paste thousands together and just put at only ONE target a cnc gen variable as money, and at all the outputs actions, like moving units it will finily cause that the AI will get money SOOOOO ****ing efficient that you wont ever beat it (note, that is how our brain works too)

 

The only difficult part are all the formula`s, which although can be found on certain pages

[Koen 2]

Neural nets are easy to make.. really...

 

This is how they work:

(snip)

 

The only difficult part are all the formula`s, which although can be found on certain pages

 

Having some experience in this field, I can tell you that the main problem is to train this 'neural net' with 100000 nodes. Even if it is a feed-forward only network with just 200 nodes, it already takes 20 minutes to train on a P4 2.4 Ghz. Neural nets are no magic trick which work on everything. You can make everything with them, but mostly they are no longer efficient enough to represent anything complex like a Generals AI.

One example of something neural nets have difficulty doing: counting. They need a few dozen units to make a simple counter.

[Guest Disposable Hero]

I was just looking for a way to simply observe 2 AI players playing against each other.

 

This would be good to check out maps (bottle necks etc.) and how good the AI scripting is.

 

Thanks for the replies!

 

Simple Just add a third player start position on the map edge(orange border), near one of the other two starting positions but don't add inner/outerperimeters for this player ! Start up the game - assign the AI players to the normal start points and team up with the AI player closest to you

 

Remember:

- don't put the 3rd starting pos. to close to one of the other because this will interfere with the AI base building

- don't put the 3rd starting pos. too far away (eg in the middle) because this will trigger the AI enemy to attack you

[Guest bws]

Neural nets are easy to make.. really...

 

This is how they work:

(snip)

 

The only difficult part are all the formula`s, which although can be found on certain pages

 

Having some experience in this field, I can tell you that the main problem is to train this 'neural net' with 100000 nodes. Even if it is a feed-forward only network with just 200 nodes, it already takes 20 minutes to train on a P4 2.4 Ghz. Neural nets are no magic trick which work on everything.

You can make everything with them, but mostly they are no longer efficient enough to represent anything complex like a Generals AI.

One example of something neural nets have difficulty doing: counting. They need a few dozen units to make a simple counter.

 

Well.. a simple generals would have about.. mm 500 nodes at max with some extra clusters to handle input and output and some clusters connected to memmory banks...

Training would be done by the creator.. just for basic generals tactics

 

When playing against it it will learn

[Koen 2]

Well.. a simple generals would have about.. mm 500 nodes at max with some extra clusters to handle input and output and some clusters connected to memmory banks...

Training would be done by the creator.. just for basic generals tactics

 

When playing against it it will learn

 

Well ok, but did I mention training is really slow? It took 20,000 cycles (games) to train a 100 node network. So you'd better not play all those games by yourself

[Guest bws]

well... they should already be trained by then... saving the current state in a file (the bias`es i thought they were called)