Page Actions
Wiki Actions
User Actions
Submit This Story

Objective Function Learning

RMSE for different approaches

  • Neural Network: 0.351
  • Linear Regression: 0.233
  • Rule Based: 0.268


TODO:

  • Reduce the feature set, e.g. remove armor and damage
  • Consider additional spatial features


DataSet

Neural Network

Configuration

  • Input layer function: linear
  • Output layer function: sigmoid
  • Hidden layer function: sigmoid
  • training cycles: 50000
  • learning rate: 0.01

Linear Regression

Configuration

  • Ridge: 1.0E-8
  • Min standardized coefficent: 1.5


Parameters

  • - 0.059 * numAllies
  • - 0.055 * numEnemies
  • + 0.981 * health
  • + 0.066 * armor
  • - 0.223 * basicDamage
  • + 0.285 * attackPotential
  • + 0.235 * distance
  • - 2.179 * enemyHealth
  • - 0.120 * enemyArmor
  • - 0.043 * enemyBasicDamage
  • - 0.238 * enemyAttackPotential
  • + 0.235 * enemyDistance
  • - 0.272 * healthRoot
  • + 0.441 * basicDamageRoot
  • - 0.111 * attackPotentialRoot
  • + 0.436 * distanceRoot
  • + 0.690 * enemyHealthRoot
  • - 0.065 * enemyBasicDamageRoot
  • + 0.155 * enemyAttackPotentialRoot
  • - 0.607 * enemyDistanceRoot
  • + 0.364 * healthSquare
  • - 0.147 * attackPotentialSquare
  • + 1.138 * distanceSquare
  • + 0.940 * enemyHealthSquare
  • + 0.023 * enemyBasicDamageSquare
  • + 0.091 * enemyAttackPotentialSquare
  • - 1.803 * enemyDistanceSquare

Rule Based

Configuration

  • Player Health Factor: 1.1
  • Enemy Health Factor: -1.1
 
linear_regression_for_objective_function.txt · Last modified: 2008/05/12 17:30 by bweber     Back to top