CITS4404_Project

Notes

• Using C++11

Sections

• Evolutionary Algorithm
• Pong Simulation
  • Done
  • Tested
• Neural Net Evaluator
  • Done
  • Partly tested

Evolutionary Algorithm

• Should handle real-valued genomes represented as vector
• Genome length must not change
• Recommend giving each individual a unique integer ID
  • Allows representation of population as map<int, vector> (map from id to genome)
• Be aware of future possibility of adding logging (so we can reconstruct the evolution)
• Beware of growing NN weights to be too large. Investigate appropriate bounds on gene values.

Pong Simulation

• Arbitrary distance units, velocities in units/tick
• Recommend mental image of 60 ticks/second to make animation (if we do it) easy later
• Construct a game instance
• Register player controller objects (one per player)
  • These will later contain the Neural Net evaluation
  • Must have a .tick() method that will be run after each simulation tick
    • On execution, reads game state, runs NN and returns desired velocity
  • Must normalize input values to between -1.0 and 1.0 for NN, and convert desired velocity produced by NN to units/tick
    • Recommend normalize such that -1.0 = far left of rectangular field and 1.0 = far right
    • Hence, vertical range is -0.75 to 0.75 or something
  • Rotate field for second player
    • Recommend coordinates from field centre, then you can just negate all coordinate/velocity values
• Remember we want to add paddle velocity to ball velocity on bounce
• First to 11 or something like that

Neural Net

• Weights represented as single vector
  • For each layer, for each neuron, for each edge to previous layer, all concatenated
  • Last weight for each neuron is to a "virtual" neuron with value 1.0
• Layers fixed at 8-8-4-2-1, given as vector
• Constructor will likely take vectors representing layers and weights and then have an .eval(vector inputs) function