Subject
AI Experiments - Possible?

Text
: A genetic alggorithm could be used, nevertheless it something really hard to do. : : I've read in the docs for some AIs that they can write new *.ini : : : : files if necessary. So I was wondering, if an AI can write a file, : : : : then can it use that ability to record its past performance and then : : : : learn from it? Here's what I have in mind: : : : : : : : : 1) In the AI's *.ini file, create a measurement request. E.g. ask : : : : the AI to write (to a file like MEASURE.DAT) the year that it : : : : finishes researching Gunpowder. : : : : : : : : 2) In the AI's *.ini file, give the AI a parameter to experiment : : : : with. E.g. the research priority that the AI assigns to : : : : Gunpowder. : : : : : : : : With both 1) and 2), the relevant line in the *.ini file might : : : : look like this: : : : : : : : : Gunpowder=5-9; measure=year @ Gunpowder=researched : : : : : : : : 3) Let the AI play many games in AI tournaments. In each game, have : : : : the AI randomly pick a research priority value for Gunpowder : : : : between 5 and 9. Also, have the AI write the year that it : : : : finished researching Gunpowder to MEASURE.DAT (along with the : : : : research priority that the AI used during that game). : : : : : : : : E.g. after a few games, MEASURE.DAT might look like this. : : : : : : : : Gunpowder=5; year_researched=400 : : : : Gunpowder=5; year_researched=360 : : : : Gunpowder=7; year_researched=-50 : : : : Gunpowder=8; year_researched=-75 : : : : Gunpowder=6; year_researched=200 : : : : Gunpowder=7; year_researched=-75 : : : : Gunpowder=9; year_researched=-100 : : : : Gunpowder=8; year_researched=0 : : : : Gunpowder=6; year_researched=280 : : : : Gunpowder=8; year_researched=25 : : : : Gunpowder=9; year_researched=-50 : : : : : : : : 4) After a suitable number of games, have the AI read the file and : : : : choose the best value of research priority based on the "year : : : : researched" data in the file. : : : : : : : : This is an obvious example, of course - because, intuitively, we : : : : *know* that the higher the research priority for Gunpowder, the : : : : earlier the AI is likely to finish researching it. : : : : : : : : But consider other possible measurements. E.g. measure how many : : : : nations an AI makes contact with by 500AD or how many techs it gets : : : : via diplomacy by 1500AD. And while measuring those, vary the : : : : priority that the AI uses in deciding whether to seek contact with : : : : other nations. Or vary the AI's research priority for techs like : : : : Map Making, Horseback Riding, or Navigation. : : : : : : : : The possible combinations of variables and measures are practic- : : : : ally endless. But with some focused experimentation, a technique : : : : like this might lead to some interesting combinations of AI param- : : : : eter settings. : : : : : : : : Does this seem feasible? : : : : : : : : : : : :

Administration

Password