Evolutionary algorithms for artificial intelligence agent learning in general video game playing

Videoigre so elektronske igre, ki z uporabnikovo pomočjo na zaslonu pokažejo vizualno povratno informacijo izbranih potez. Njihov osnovni namen je zabava in krajšanje časa. V zadnjih petih letih se je z mednarodnim tekovanjem inteligentnih agentov za igranje iger (angl. General Video Game AI competitionv nadaljevanju GVGAI) začelo novo poglavje. Tekmovanje GVGAI od udeležencev zahteva stvaritev agenta, ki s pomočjo optimizacijskih algoritmov poskuša doseči najboljši možen rezultat. Ker se nam je tekmovanje GVGAI zdelo zelo zanimivo, smo se odločili ustvariti agenta, ki s pomočjo evolucijskih algoritmov pri igranju videoiger, doseže kar se da dober rezultat. Agenta smo zasnovali po pregledu obstoječih optimizacijskih algoritmov. Za razliko od ostalih agentov, naš agent uporablja diferencialno evolucijo, ki še ni bila prikazana na tekmovanjih GVGAI. Dobljene rezultate primerjamo s pomočjo primerjalnega preizkusa GVGAI, vidimo pa da je naš agent statistično signifikantno boljši od večine, a obstaja prostor za napredek.Video games are electronic games that show us visual feedback on the screen, based on the actions selected by the user. Their basic purpose is fun and entretainment. In the last five years, a new chapter for video gaming has opened in the form of GVGAI competition. The competition challanges the contestant to implement an agent that can maximize the score of played video games with usage of modern optimization algorithms. To us, the idea seemed very intriguing, so we decided to implement an agent that relies on evolutionary algorithms and achieves the highes score possible. We designed our agent after reviewing the existing optimization algorithms. Our agent uses diferential evolution, which was not yet used in a GVGAI competition. Our results are compared using the GVGAI benchmark and as we can see from the results our agent is statistically significantly better than most of the existing ones, but there is still room for improvement