The Learning Rate

• A high learning rate is useful when learning begins, as very little information is encoded in the Q-estimates
• As learning continues, more and more information is encoded in the estimates
• Consequently, it is worthwhile to start with a high learning rate that decreases with time
• Example
  • α = 1 / (1 + (t/c))
  • t is the current timestep
  • c is a fixed constant