Neuroevolution and High Dimensional Datasets

1
Neuroevolution and High Dimensional Datasets

• John Hettinger
• cse848
• 5 December 2006

2
Evolving Motivations

• Original Motives
• Classify natural language texts
• Analyze augmenting topologies
• Determine usefulness of recurrent networks

3
Brick Wall Reorientation

• Current Motives
• Survey augmenting topology systems
• Analyze complexity of NEAT
• Evolve a successful NN classifier (dataset with a
  moderate number of dimensions)

4
Hypotheses

• NEAT scales up well
• Recurrent networks are effective efficient
• Evolution can be helpful with natural language
  processing (classification)

5
Project Design

• Original Framework
• Compile dataset
• Write custom NEAT functions
• Evaluation function uses training dataset
• Test dataset used to measure winning individuals
• Analyze resulting networks and evolutionary
  histories
• Run under various conditions

6
Problems and Lessons

• Focus on priorities
• NEAT does not scale gracefully
• Start small (time and/or space)
• The usual (what did you tell the GA to do?)
• Classification and GAs (input/output)
• Testing and over-fitting

7
Conclusions

• NEAT excels at low dimensional reinforcement
  learning
• Evaluation function is crucial sensitive
• Pattern recognition tasks should use traditional
  techniques to verify efficacy
• Recurrent networks are irrelevant when solving
  toy problems
• Complexification easily gets out of control

8
Token Graphic (see blackboard)

• / Organism 52 Fitness 10.6825 Error 0.731584
  /
• / ------ WINNER 52 SPECIES 36 ------ /
• genomestart 52
• trait 1 0.1 0 0 0 0 0 0 0
• trait 2 0.523617 0.350709 0.0287805 0 0 0.749028
  0.854966 0
• trait 3 0.3 0 0 0 0 0 0 0
• node 1 2 1 3
• node 2 3 1 1
• node 3 1 1 1
• node 4 1 0 2
• node 10 1 0 0
• node 21 1 0 0
• node 123 1 0 0
• gene 2 1 4 -1.13573 0 1 -1.13573 1
• gene 2 2 4 -0.473696 0 2 -0.473696 1
• gene 3 3 4 -0.504511 0 3 -0.504511 0
• gene 3 3 10 2.83022 0 14 2.83022 1
• gene 3 10 4 2.7588 0 15 2.7588 1
• gene 3 3 21 -2.51245 0 43 -2.51245 1

9
New Hypotheses (future work)

• Can divide and conquer work and tame high
  dimensional inputs?
• Is replacement of some O(n2) code enough for
  acceptable performance?
• How important is space complexity? (bitmaps and
  other storage efficiencies)

10
References Resources

• Evolving Neural Networks through Augmenting
  Topologies by Kenneth O. Stanley and Risto
  Miikkulainen, Evolutionary Computation, Vol. 10,
  No. 2, pp. 99-127, 2002
• Real-Time Neuroevolution in the NERO Video Game
  by Kenneth O. Stanley, Bobby D. Bryant, Student
  Member, IEEE, and Risto Miikkulainen, IEEE
  Transactions on Evolutionary Computation, Vol. 9,
  No. 6, December 2005
• http//www.cs.ucf.edu/kstanley/software