Forum Overview :: Are Games Art? :: Re: You just had to screw it up in the thread where you prove Monty wrong

[quote name="mark"][quote name="Entropy Stew"] GA != intelligence, and I don't know why people keep saying it does. Does the evolutionary process GA is based off of think? Fuck no. All it is is directed semi-random meanderings across the entire problem domain. [/quote] Who (other than SB, I guess) says it equals intelligence? It would be pretty impressive that intelligence could be contained in tiny amount of code + a random number generator. If this is all bottom up means, then yes logicians would not be interested in it. Determinism has its place and, frankly, GA are sort of boring. [quote] [quote]The trick is to use heuristic maps to set ranges within which highly flexible neural networks can run wild, using trial and error to teach them what works and what does not on a visual level.[/quote] That's even stupider than your Carmack-as-the-competent-implementor platform. 1. Your solution is actually a horribly disfigured genetic algorithm [/quote] Right and disfigured means 'unlikely to yield correct answer.' Here 'correct answer' is unclear because we (as you point out below) have no metric top judge our success. Even if we did, GA are proven to converge on the right answer as time goes to infinity because it tries every stupid solution. In a system where we set ranges, we claim to already know so much about the problem space that maybe we should be using deterministic methods. Probably we aren't sure, so we should not set ranges at all, other than to eliminate values that give impossible results (something, again, probably impossible in the scope of auto-generating a videogame). [quote] 2. It would (sort of) work if the fitness evaluation function could JUDGE THE QUALITY OF THE GAMEWORLD 3. Judging the quality of the gameworld is AI-complete 4. The code to generate the world for each iteration would be the largest thing ever written if you wanted any sort of originality. 100% originality for this process would also be AI-complete. [/quote] I am suspicious of a term like AI-complete. With NP-complete problems, we have a number of nice qualities among these: a reduction proof which proves NP-completeness, and the ability to understand how to prehaps reduce the complexity of a problem until it winds up in P. On the general class of graphs Longest Path is NP-complete, it is in P on a DAG; can any AI-complete be provably reduced to a computable form? Regardless, I agree that the above problems are very very hard. Maybe on some level people are too quick to equate AI-Complete with NP-Complete which is why we end up with hill-climbing as the ultimate solution. mark [/quote]