It’s
rather unfortunate that I couldn’t be there for the Mind Matters conference,
and that I’m missing this wonderfully-themed Cognitive Séance as
well. Even so, I’ll contribute by adding a few thoughts. I started typing this
stuff out as a wall post, but my words have once again outgrown their container,
so I’ve moved them to my blog. Well, here’s what I have:
A couple
of useful approximations of wisdom that I’ve found useful are “problem solving
ability in the absence of expertise”, and “meta-expertise”. Neither is perfect,
but both are instructive. The first can be understood as being
“execution-level”, while the second is “development-level”. The performance-competence
distinction comes to mind.
Goals and
problems are relevant to discussions of wisdom. Problem detection and goal
selection are flip-sides of the same coin. The first indicates an external
disruption of homeostasis, while the second suggests an internal redefinition
of the same. The agitation or dissonance that this state of affairs produces is
what moves the agent to act.
Action
can be external or internal. External action, as Searle might say, attempts to
fit the world to the agent’s preferences, while internal action attempts to fit
the agent’s preferences to the world. So long as the world-mind difference is
reduced sufficiently, the “problem” will have been solved satisfactorily.
To
witness wisdom in action typically requires an unfamiliar state of affairs. One
kind of foolishness is to mistake the unfamiliar-yet-irrelevant as a problem.
If the unfamiliar scenario is genuinely unacceptable, though, a wiser agent
(understood in some manner that is independent of this problem) will be more
likely to find a solution than a less wise agent (assuming that neither have
domain expertise).
Since
expertise is out of the picture, wise agents appear to rely on heuristics. If
two agents have identical repertoires of heuristics, the one that selects a subset
for use in a more context-sensitive fashion is more likely to reach a solution.
Having a larger repertoire of heuristics, I would imagine, might initially
provide an advantage, but would eventually lead to inefficient heuristic
selection.
These
execution-level thoughts about wisdom apply “at runtime”, so to speak. On the
other hand, wisdom also involves development-level components. To continue the
analogy, they are “compile-time” factors. This is where “wisdom
as
meta-expertise” enters the discussion. For
example, the heuristics an agent learns over time affects how well it can solve
an unfamiliar problem. This means that some deviation from homeostasis
encouraged the agent to revise its repertoire of heuristics. It (partially)
solved its “sub-optimal heuristics” problem, and if it practises this skill
enough, it might gain expertise in heuristics optimization. Similarly,
acquiring expertise various specific domains might lead to “expertise” in
expertise acquisition. The reason I’ve used scare-quotes is that
“domain-general expertise” is an oxymoron, though I’m not sure whether “expert
generalizer” is.
Now I’ll
turn to Artificial Wisdom (AW). Some artificial expert systems that exist today
have been spoon-fed their abilities, while others have become experts through
practise. It’s this second type of artificial systems that are more likely to
act as the forerunners to AW. One important tool that helps humans acquire
expertise is generalizing across different instances of the same kind of
problem. “The star-shaped peg fits in the star-shaped hole” and other blindly-discovered
solutions to similar problems can lead an intelligent agent to heuristics like
“look for a hole that’s the same shape as this peg”.
Notice
how I’ve performed a bit of recursion above. Generalizing across instances of
the “acquire expertise” problem, I’ve picked out the “generalize over problem
instances” heuristic. Continuing this recursive trend might prove fruitful to
AW research.
Wise
agents probably excel in identifying accidental successes and developing them
into reliable algorithms and heuristics, kind of how Jeff Hawkins describes the
neuroscience of rehearsal in On
Intelligence. Other essential mechanisms for developing AW include ways to
evaluate similarity and relevance.
I hope
all of you have a great Séance!
