In the book "Reinforcement Learning: An Introduction" (2018) Sutton and Barto explain at page 221 a form of tile coding using hashing, to reduce memory consumption.
I have two questions about that:
- How can this approach reduce memory consumption? Isn't it just dependent on the number of tiles (you have to store one weight for each tile).
- They state that there is only a "little loss of performance". In my understanding, the sense of tile coding (and coarse coding) is, that near-by states have many tiles in common and far-away states have only few tilings in common. With tilings "randomly spread throughout the state space" this isn't the case. How does this not influence performance?