Multi-Objective Neuroevolution in Super Mario Bros.
This thesis explores how to use Multi-Objective Evolutionary Algorithms (MOEA)
to solve problems that are not explicitly defined as multi-objective problems. A
neuroevolution technique consisting of combining a multi-objective evolutionary
algorithm called NSGA-II and artificial neural networks (ANN) based on Neu-
roEvolution of Augmented Topoligies (NEAT) were used to develop a system
that created controllers for a version of the Super Mario Bros game called Mario
AI. Experiments were conducted to measure different ways to define objectives
for MOEAs in Mario AI, how using these objectives as a basis for a scalar fitness
function would affect a genetic algorithm and to examine how to use ensembles
to combine individuals of a pareto front into a single controller that would be
able to display the strengths of all of the individual controllers.
The results show that adding sub-goals as objectives together with the main goal
could have a positive effect for a MOEA and that the same sub-goals could also
give a positive effect when applied to the scalar fitness of a genetic algorithm.
It is however not trivial to decide which sub-goals to use, as most of the chosen
objectives were found to have a negative impact on the controllers, even when
selected based on the authors? expert knowledge about the game domain. Using
basic behaviours that the controller has to use in order to play well as objectives
had a negative effect on the controllers and the controllers were able to learn
these behaviors even without using them as objectives.
Place, publisher, year, edition, pages
Institutt for datateknikk og informasjonsvitenskap , 2013. , 82 p.
IdentifiersURN: urn:nbn:no:ntnu:diva-23600Local ID: ntnudaim:9651OAI: oai:DiVA.org:ntnu-23600DiVA: diva2:676807
Haddow, Pauline, Førsteamanuensis