Cooperative, Competitive and Mixed cooperative-competitive multi-agent reinforcement learning for the game Snake

Zadatak i cilj rada je opisati problem pronalaženja i optimizacije strategije igranja u igrama s više igrača (agenata) te realizirati model učenja strategije i sustav za učenje temeljen na evolucijskim algoritmima. Razvijeno je nekoliko verzija igre "Zmija" koje se razlikuju ovisno o cilju više agenata za igranje. Kroz svaki od kojih se modelira optimizacija. Optimizacija je modelirana kao koevolucija u evolucijskom algoritmu dok su agenti genetski programi temeljeni na stablu. Eksperimentalno se kroz rad utvrđuje učinkovitost nekoliko optimizacijskih algoritama ovisnih o parametrima okruženja za učenje. Kao glavni konkurenti nekoliko proučavanih optimizacijskih algoritama, razmatraju se modifikacijske tehnike implementiranog algoritma učenja – Joint-action učenje i Self-play. U radu je prikazan izvorni kod, dobiveni rezultati uz njihovo objašnjenje, razradu i korištenu literaturu.The task and goal of the dissertation is to describe the problem of finding and optimizing a playing strategy in games with multiple players (agents) and to realize a strategy learning model and system for learning based on evolutionary algorithms. Several versions of the game "Snake" are developed that vary by the goal of the multiple playing agents through each of which optimization is modeled. The optimization is modeled as a coevolution in an evolutionary algorithm while the agents are tree-based genetic programs. Experimentally, throughout the dissertation, the effectiveness of several optimization algorithms dependent on the learning environment parameters is determined. As top contenders of the several studied optimization algorithms, the modification techniques of the implemented learning algorithm joint-action learning and self-play are considered. The dissertation presents the source code, the obtained results alongside their explanation, elaboration, and utilized literature