[e-lang] Decision Alignment (was: E research topics)
Mark S. Miller
markm at cs.jhu.edu
Sun Apr 15 00:39:58 CDT 2007
More thoughts on research topics.
Decision Alignment <http://erights.org/decision/>.
Bill Tulloh and I have been talking about a larger taxonomic framework in
which to organize plan coordination issues, both the emphasize the analogies
between the human world and the computational world, as well as to draw
attention to the continuity of these issues across the interface between these
Say we have a system complex enough to describe in terms of interacting
intensional entities, say Alice and Bob, and potential cooperative
interactions between Alice and Bob. Imagine that Alice is attempting to
subcontract/delegate some portion of her job to Bob. In any such system, we
need to examine how Alice can influence Bob to behave in a manner more likely
to serve Alice's interests.
In the human world, when the question is put this way, the conventional
economic point of view too quickly focuses on incentives almost exclusively as
the answer. Indeed, our overall label, "Decision Alignment", is a takeoff on
I say "too quick" because incentives only become the limiting issue after
various more fundamental issues have been dealt with. For example, how hard is
it for Alice to explain to Bob what Alice needs in terms Bob can understand
and act on? By jumping to incentive issues, the conventional economic
perspective assume away the difficulty of these logically prior issues.
Even after explanations are adequate, incentives are rarely sufficient by
themselves for Alice to shape Bob's behavior, because Bob may have too many
"moral hazard" opportunities. By contrast, if Alice can grant Bob only narrow
least authority, which is as relevant in the human world as we hope it will be
in the computational world, this will often reduce Alice's risks from Bob more
than incentive issues can.
In the human world, the distinction between what we call "inspection" vs
"monitoring" is often fuzzy. But the logical distinction is still meaningful.
By "monitoring", we mean watching (sampling) Bob's output, how he behaves. By
"inspection", we mean examining Bob's internal mechanism, how he's
constituted, to figure out if the logic of Bob's internal mechanism means he's
unlikely to misbehave in certain ways.
Once we see that Alice can use all five techniques to shape Bob's behavior, we
can examine how Alice can make synergy between these techniques. For example,
when Alice couples least authority with incentive alignment, now Alice only
has to worry about moral hazards within the narrower set of choices left open
Text by me above is hereby placed in the public domain
More information about the e-lang