Decentralized Markets versus Central Control: A Comparative Study

This study takes previous work claiming the superiority of decentralized control for a building climate control system compared to a previous central implementation. it knocks this claim down, building a centralized system that is better. It than retakes the other papers claim, by making a decentralized controller that is even better.

The argument for why the first decentral controller was better than its control was that the decentral one had complete information abou tthe system. Once you give a centralized controller complete information, its performance can be optimal.

They then implemented a decentralized controller whose behavior is also optimal.

This paper is definitely coming from the direction of game theory and optimal control theory (Pareto efficient, globally optimal, utility functions), which assumes (and picks problem that have) no limits on computing power/time and complete information. Given these assumptions, a centralized with all ways be better. The best a distributed approach can hope for is to mimic the behavior of the central optimal approach (which is what theirs does).

Not sure why, but when folks from the optimal control school get into multi agent systems, it is always markets. Must be that game theory has proofs. But presumably it doesn't matter: " It is tempting to ask whether things are differetn wheen a non- market multi-agent approach is followed , say, using the contract net. As we argued, the answer in our opinion is a stright-forward no. The goal in the condiered class of problems is to find the optimal distributed solution. Alternative agent approaches, market as well as non-market ones, only change the multi-0agent dynamics on the way to this goal. this might be doen in a better or poorer way, but it is not possible to change this goal itself. "

also: "as a conjecture we state the following general result. Multi-agent equilibruim markets yield otpimal decentralized solutions to distributed resource allocation problems that are of the same quality, in terms of a given overall systems performace index, as solutions fiven by the optimal ceent central systems controller that has acess to all relecant loacl data. ..."

I think this all assumes independance of the agents. Phrasing of this argument has to be careful because one mistep and the math behind optimal control theory becomes intractable for large systems.

"... this is a srong and general statement about the relationship(and evenoutcomeequibalence) between decentralized markets and central control" --I think where he is taking that is that agents actually are to be preferred in certain cases, and may be a way of attaining optimal control in systems to large or complicated for a central controller.

They really tear apart the Huberman and Clearwater, showing that many of the parameters central to their model were entirely irrelevant (essentially that they took the market metaphor way too literally).

Ooh, the paper has proofs, i forgot about that.