Formal Cartoon for AI KR & Planning

[josh]

What kind of formalisms help to structure computation?

For reasoning & planning we want to be able to consider lattices as graphs at different levels of abstraction, where nodes are combinations of environment/situation/episode & paths represent change with different costs in terms of possibility of occurrence, effort, cost in resources, cost in time, uncertainty vs. value of exploration for kd, etc.

In most problems we focus first on high levels of abstraction including nodes that are believed to be easily accessible and available as “low cost” paths from where we are now.

Graph edit has different kinds of operations – data update, theory update, refinement, etc.

We can look for efficient models in VR that perform well on all of the VR training data (modulo usual ML issues of variance/bias tradeoff, cross validation/bootstrap, simplicity priors, etc.)

Keep building around the viable paths toward more complexity/mess meeting greater ranges of capabilities.

[josh]

My personal gestalt about KR wants to emphasize that there is real gain to be had in the act of getting our research thought patterns away from being stuck in “which set of symbolic predicate/argument is here”. Why? The first course one takes in predicate/first order logic/proof is massively enlightening. Everyone should do that as soon as they are ready. Formal theories map to 1st order logic or higher order when special infinite sequences obtain. Why deviate in form?

Why is:

T1) he role of environment & environmental variation is ignored in the most of the NL like discourse of analytical philosophy by habit.

2) VR covers don’t map nicely to NL.

3) Time & spatial scale usually matter when they are large and are always available for change – predication isn’t helping with boundaries; that’s where vagueness enters

4) Lots of causally relevant detail doesn’t have ez predicate representation available in a NL like form

5) Reasoning about situations is always partial & the partiality is not based on necessity or sufficiency when environments & other factors are considered.

6) If you look carefully, what is the predicat & what is the argment can usually be flipped in most natural language expression – so the form is already not very constraining in simple sentences.

I’m championing the position that the right theory for knowledge engineering is going to be unlike a variant of first order logic, and the key difference is not about making all the fields floats or doubles or arbitrary precision numbers. So let’s look for a better cartoon theory of mind to support the frontiers of knowledge engineering today.

[josh]

The graph framework described above can also lend itself to search/planning/exploration of maker/synthesis/process frameworks where environments are directly controlled & there is less ambiguity about terminology but higher levels of demand for precise specification of measurement related properties.

[josh]

In other words, alternative processes correspond to nearby nodes in the process design space.

[Author]

Posts