Rules and Rule Learning

In some sense you can see how a rule is just a simple if conditional.
If CA1 and CA2 are on the do CA3.
However, for a bit more sophistication, variable binding is needed.
A standard grammatical rule is NP + VP -> VP and the NP is the actor (I saw).
This requires at least two variables (one for the NP and one for the VP).
We did a simple form of this with our counting paper back at ICCM in 2006 but we used binding via LTP.
The first stack based parser from CABot does the later task.
So, we can do rules in this system.
However, as in processing, the real issue is how does the system learn rules.
We presented a paper at ECAI this summer about a system that learns goal action pairs.
It makes use of environmental feedback in a kind of semi-supervised learning.
So, a goal is set.
The system then selects an action.
If the action results in the goal being fulfilled, the system reinforces the action.
If not it selects another action.
This works but has problem with the bucket brigade.
It is a neural implementation of my colleague Roman Belavkin's theoretical work on optimal learning. To some extent this solves the exploration exploitation dilemma.
We also need to figure out how to incorporate variables into the learning.