Improved Learning
- Before we did the CABots, I worked on learning.
- I introduced a compensatory learning rule (Steve Kaplan gave it to
me). Neurons have a total amount of synaptic strength they can
allocate.
- I also did some work with learnin Hierarchical categories.
- Aside from some variable binding work, CABot1, 2 and 3, have
very little learning.
- CABot4 has a bit more, and CABot5 has a lot.
- We'll work on learning procedures,
- learning goal action pairs (following the ACT and Soar work),
- learning rules with variable bindings (e.g. syntax),
- learning new CAs appropriately (CA3 problem),
- learning associations,
- learning more complex structures (like cognitive maps).
- A range of approaches may be tried, but reinforcement from
environmental feedback is the action pairs method (Roman
Belavkin).
