Feedback 26-07-12

Progress

• Thesis: 
• Wrote a section about related work
• Started to write on conclusion chapter
  
• Added experiments about performance in infinite horizon tiger problem

Overview of experiments

• Performance
• Sample based (# roll-outs vs mean)
1. Variations of COMCTS
2. Variations of TT
3. Best COMCTS variation against best TT variation
4. Variations with similar performance
• Time based: same four experiments (time vs mean)
• Practical time and space usage (time vs roll-outs)
• Varying Noise (Standard deviation of noise vs mean)

Experiments: Time vs roll-outs, time vs mean

Set-up

• Environment: infinite horizon Tiger problem
• Discount factor: 0.5
• Discount horizon: 0.00001
• Number of episodes: 2,000
• Algorithms:
• BFAIRMC (pronounced: "be fair mc") = Belief based Function Approximation using Incremental Regression trees and Monte Carlo = transposition tree

Results

time vs rollouts

time vs mean

time vs mean (zoomed in)

time vs regret (optimum - mean)