Evolving an AI for Spring

[Jonanin]

Zerg, anyone?

Or wait.. does the Spring community hate Starcraft?

[Nemo]

I was just pointing out that the particular strategy of flashspam+rocket trucks (which extends well beyond the "rush" stage and into the very late game) took literally years before it turned into the all-powerful strategy it has been in competitive OTA for a while now. So I wouldn't expect an AI to arrive at it particularly quickly.

[smokingwreckage]

I'm a TA fan and I like Starcraft. Maybe that makes me a sicko.

[Dragon45]

Goddamn it, why did I see this discussion so late:

Short Answer: GA, ANN, or most 'biologically inspired optimization methods' are wonderful to think about, but a MESS to implement IRL. Too many iterations, too much processing power. In short, far too resource-intensive.

As 'cool' as biologically inspired optimization is, remember that at the end of the day we already have the best optimizers in the world that have conducted tens of thousands of iterations for us already - human beings. Human beings in this case are the GA; we have a plethora of effective strategies that have been formulated that have been shown to be the 'best'. The best GA we can ever work with is Randy, Steven, and Lion. Copy them, use their strategies as the overarching framework in an AI. They work. They've been proven to work. It's far more feasible to code manually the majority of cases instead of wanting to 'evolve' it.

People sometimes forget that straight optimization is often the best way. Unless you're doing it 'just to see what will happen' (which is a perfectly valid justification tbh, i do this shit myself); just remember that perhaps only Google has the processing power to actually crunch the numbers for this one evolutionarily speaking.

[eriatarka]

Short Answer: GA, ANN, or most 'biologically inspired optimization methods' are wonderful to think about, but a MESS to implement IRL. Too many iterations, too much processing power. In short, far too resource-intensive.

As an aside, ANNs have nice applications in more concisely defined classification problems, and GAs are used on some classes of optimization problems, IIRC. So they're not all useless. But I agree that using this kind of methods to 'evolve' an AI is completely infeasible for a game of this complexity.

[pharoph]

something i just wanted to point out... when im playing spring with KAI i like to kill his commander. but when i do all the factorys stop producing units. go fix dog.

[Forboding Angel]

/stab
/stab
/stab