Optimization, Imitation and Innovation: Computational Intelligence and Games

Created by W.Langdon from gp-bibliography.bib Revision:1.4208

  author =       "Julian Togelius",
  title =        "Optimization, Imitation and Innovation: Computational
                 Intelligence and Games",
  school =       "Department of Computer Science, University of Essex",
  year =         "2007",
  month =        sep,
  keywords =     "genetic algorithms, genetic programming",
  URL =          "http://julian.togelius.com/thesis.pdf",
  size =         "202 pages",
  abstract =     "This thesis concerns the application of computational
                 intelligence techniques, mainly neural networks and
                 evolutionary computation, to computer games. This
                 research has three parallel and non-exclusive goals: to
                 develop ways of testing machine learning algorithms, to
                 augment the entertainment value of computer games, and
                 to study the conditions under which complex general
                 intelligence can evolve. Each of these goals is
                 discussed at some length, and the research described is
                 also discussed in the light of current open questions
                 in computational intelligence in general and
                 evolutionary robotics in particular.

                 A number of experiments are presented, divided into
                 three chapters: optimisation, imitation and innovation.
                 The experiments in the optimization chapter deals with
                 optimising certain aspects of computer games using
                 unambiguous fitness measures and evolutionary
                 algorithms or other reinforcement learning algorithms.
                 In the imitation chapter, supervised learning
                 techniques are used to imitate aspects of behaviour or
                 dynamics. Finally, the innovation chapter provides
                 examples of using evolutionary algorithms not as pure
                 optimisers, but rather as innovating new behaviour or
                 structures using complex, nontrivial fitness

                 Most of the experiments in this thesis are performed in
                 one of two games based on a simple car racing
                 simulator, and one of the experiments extends this
                 simulator to the control of a real world
                 radio-controlled model car. The other games that are
                 used as experimental environments are a helicopter
                 simulation game and the multi-agent foraging game

                 Among the main achievements of the thesis are a method
                 for personalised content creation based on modelling
                 player behaviour and evolving new game content (such as
                 racing tracks), a method for evolving control for
                 non-recoverable robots (such as racing cars) using
                 multiple models, and a method for multi-population
                 competitive co-evolution.",

Genetic Programming entries for Julian Togelius