Ex: An Effective Algorithm for Continuous Actions Reinforcement Learning Problems. Martin H., de Lope

• The algorithm that won the 2009 RL competition in the helicopter domain (I believe)
• Uses a weighted k-nn approach along with TD methods to estimate value across continuous actions
  • Uses “probability traces” instead of eligibility traces, could be more accurately be called “weighted eligibility traces”
  • In general they use the word “probability” when “weights” seems more accurate
• Doesn’t have any proofs, but is empirically effective.  Seems like a reasonable approach
• Database of Q-values grows as more samples are taken, so more points to interpolate between with weights as more experience occurs
• Experimental section would be better if it averaged the results of a number of experiments as opposed to just one run of an experiment

Advertisements