Control by Committee

[Alexander Bistagne]

Link post

I think I have an interesting new research direction: Aligning Committees.
Make ensembles of agents and have them act together as only one agent in the world with some protocol for how they combine preferences.

Main Motivating Construction:
Given a target consequence T, we construct a committee out of 3 expected utility maximizing agents: Planner, Wanter, and Unwanter (which have different utility functions). Planner proposes a plan and then Wanter and Unwanter either signoff or veto said plan. A plan P which gets both to sign off is executed, otherwise the agent does the safe but useless null action $\varnothing .$

The Utility Functions

$U_{planner}=\{\begin{array}{cccc}1& if& P& happens\\ 0& if& \varnothing & happens\end{array}$
$U_{wanter}=\{\begin{array}{cccc}w& if& \varnothing & happens\\ 1& if& P\wedge T& happens\\ 0& if& P\wedge \neg T& happens\end{array}$

$U_{unwanter}=\{\begin{array}{cccc}1-n& if& \varnothing & happens\\ 1& if& P\wedge \neg T& happens\\ 0& if& P\wedge T& happens\end{array}$

Then as the committee gets smarter and more powerful $Pr\left[T|P\right]\in (w,n)$ I believe this is a mild optimizer.

Questions:
1. Is this already part of someone else’s research agenda? if so, who?
2. Is there anything I should definitely read before heading down this path?
3. Is the main motivating construction a mild optimizer? why or why not?

If this is interesting to you, please read the full post at the link.

[Alexander Bistagne]

New location of previous file on github: https://​​github.com/​​Alexhb61/​​Alignment/​​blob/​​main/​​committees/​​Control_by_committee.md

The main criticism I imagined of this piece was the disassembly risk.

By adding a latter step to this outer alignment approach (transforming a committee into a modified learning problem), I believe I have a much stronger angle on outer alignment.
I will continue to write about it here:
https://​​github.com/​​Alexhb61/​​Alignment/​​blob/​​main/​​committees/​​outer_alignment.md