CAs for Controllers
- The CABot agents had a few (e.g. 8) orthoganal CAs for actions
like left or up.
- More sophisticated controller dependent CAs would need to be developed.
- An early question is what type of effectors are there?
- There has been some work (Palm) with CAs to control a robotic
camera position.
- We've done some work with continously valued input (and even
less on output), but this should be relatively easy.
- More interesting is to learn relations between the outputs
and other parts of the system.
- Multiple controllers could also work together. (Proposal
with Tony White on aircraft controllers.)
- Another option is to have multi-layer controllers to
trade off between learning and accuracy.
