NeuroHunter — an Entry for the Balanced Diet Contest

by Wojciech Jaśkowski, Krzysztof Krawiec, Bartosz Wieloch

Abstract:

NeuroHunter is the name of an evolved neural network that won the GECCO’2008 Balanced Diet contest, organized within GECCO’2008, Genetic and Evolutionary Computation Conference, in Atlanta, Georgia. The goal was to evolve a controller for a virtual agent that collects two types of food in a 256×256 board, trying to maximize the amount of food collected while minimizing the imbalance between the two types of food. The major difficulty of the task consists in partial observability: the agent cannot see the food pieces. However, the probability of finding a food piece of particular type at a particular location depends on another feature, elevation, which is visible to the agent. Thus, the agent has to learn this dependency, also avoiding revisiting the same locations. Our winner entry used the HyperNEAT by Stanley et al., an evolved structural neural net. The report describes in detail the method, technology, and the results.

View PDF

Reference:

NeuroHunter — an Entry for the Balanced Diet Contest (Wojciech Jaśkowski, Krzysztof Krawiec, Bartosz Wieloch), Technical report, Institute of Computing Science, Poznan University of Technology, 2008. (3)

Bibtex Entry:

@TechReport{Jaskowski2008neurohunter,
  Title                    = {{NeuroHunter} --- an Entry for the Balanced Diet Contest},
  Author                   = {Wojciech Jaśkowski and Krzysztof Krawiec and Bartosz Wieloch},
  Institution              = {Institute of Computing Science, Poznan University of Technology},
  Year                     = {2008},

  Address                  = {Poznań, Poland},
  Number                   = {RA-10/08},

  Abstract                 = {NeuroHunter is the name of an evolved neural network that won the GECCO'2008 Balanced Diet contest, organized within GECCO'2008, Genetic and Evolutionary Computation Conference, in Atlanta, Georgia. The goal was to evolve a controller for a virtual agent that collects two types of food in a 256x256 board, trying to maximize the amount of food collected while minimizing the imbalance between the two types of food. The major difficulty of the task consists in partial observability: the agent cannot see the food pieces. However, the probability of finding a food piece of particular type at a particular location depends on another feature, elevation, which is visible to the agent. Thus, the agent has to learn this dependency, also avoiding revisiting the same locations. Our winner entry used the HyperNEAT by Stanley et al., an evolved structural neural net. The report describes in detail the method, technology, and the results.},
  Comment                  = {3},
  Homepage-url             = {http://www.cs.put.poznan.pl/wjaskowski},
  Keywords                 = {evolutionary computation, genetic programming, evolutionary algorithms, games, coevolution, artificial intelligence},
  Order                    = {4},
  Url                      = {http://www.cs.put.poznan.pl/wjaskowski/pub/papers/jaskowski08nerohunter.pdf}
}