Released 2014
Modular Multiobjective NEAT is a software framework in Java that builds on the basic principles of Neuro-Evolution of Augmenting Topologies. MM-NEAT uses Non-Dominated Sorting Genetic Algorithm II to carry out multiobjective evolution, and supports networks with multiple output modules. Evolved agents can use different modules for different behaviors: Human-specified task divisions can be defined using the Multitask Learning approach, or the task division can be learned using preference neurons. Evolution can also learn how many modules to use with various forms of Module Mutation. Finally, multiobjective evolution can be improved using Targeting Unachieved Goals (TUG), a fitness-based shaping technique that turns objectives off when they are not needed. For more information on these techniques, see the associated publications.

The primary domain in which these methods were evaluated is Ms. Pac-Man. Batch files are included to recreate all experiments from Jacob Schrum's 2014 dissertation. Some RL-Glue domains are also included, as well as a simplified version of BREVE Monsters, which is a precursor to MM-NEAT.
Jacob Schrum Ph.D. Alumni schrum2 [at] southwestern edu
Discovering Multimodal Behavior in Ms. Pac-Man through Evolution of Modular Neural Networks 2015
Jacob Schrum and Risto Miikkulainen, To Appear In IEEE Transactions on Computational Intelligence and AI in Games (2015).
Evolving Multimodal Behavior Through Modular Multiobjective Neuroevolution 2014
Jacob Schrum, PhD Thesis, The University of Texas at Austin. Tech Report TR-14-07.
Evolving Multimodal Behavior With Modular Neural Networks in Ms. Pac-Man 2014
Jacob Schrum and Risto Miikkulainen, In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO 2014), pp. 325--332, Vancouver, BC, Canada, July 2014. Best Paper: Digital Entertainment and Arts.