Homepage



  » Welcome
  » About
  » Hosting

  General AI



  » Introductions
  » Finite State Machines
  » Ameliorated Future
  » Miscellaneous

  Neural Networks



  » Introductions
  » Backpropagation
  » Optimization
  » Simulators
  » Applied NNs
  » Sample Code
  » Image Recognition
  » Image Compression

  Artificial Life



  » Tutorials
  » Sample Code

  Genetic Algorithms



  » Libraries
  » Sample Code

  Fuzzy Logic



  » FAQ

  Games



  » Sample Code

  Reinforcement Learning



  » Tutorials
  » FAQ

You are in: Reinforcement Learning  /  FAQ  /  Nuts and Bolts of RL  /  Why do you call it "tile coding" instead of "CMACs"?
Why do you call it "tile coding" instead of "CMACs"?

The idea of tile coding is essentially the same as that underlying CMACs. Without in any way intending to detract from the credit due the pioneers of CMACs (e.g, Albus, Marr, Miller...), sometimes it is better to switch to a new name. The name CMAC stands for, among other things, "Cerebellar Model Articulatory Controller," which seems pretty inappropriate for the current usage. The original CMAC also used a slightly different algorithm -- a correlation rule rather than an error correction rule. The use in reinforcement learning steps it even farther away from its original use. What remains is not so much a learning system (much less a cerebellar model), but a way of coding states for use by a learning system. The key features of the coding is that it is based on multiple exhaustive partitions (tilings!) of the state space, and that it is particularly suited for implementation on serial, digital computers.






Download Article
Printer Friendly
Back


All content copyrighted by Avaye.com