Monday, November 3, 2008

Further Market-based-AI ideas

A friend commented on my post about about research ideas for market mechanisms in AI. He brought up a couple of excellent points. Here's my reply, which I feel's worth posting:


Excellent question! I've not done much thinking along the lines of inter-agent efficiencies; that will definitely be fertile ground to develop. On the specific question of which agents a given agent interacts with and how those links are established: I agree, your "case 1" wouldn't allow the kind of dynamic flexibility that a market or a mind needs.

To extend the metaphor, one approach to the second case might be to expand the role of the "investor" agents. One of their functions (and methods of earning a return) could be to "fund" new "start-up" agents. There might be a process where agents are generated spontaneously, each of which proposes a set of connections and a transformation. The investor agents could evaluate the proposals and select a few to give a big boost of activation (eg: funding). The unselected ones struggle on, probably failing quickly; the selected ones get to compete on more equal footing with the existing agents.

That could be called the "entrepreneur" method of new connections, another might be the "sales call." Existing agents could propose connections with others, and they would be evaluated by the receiving agent. They might go through a process of "trial shipments," which would determine whether or not the connection lasts. The other side of this scenario would be the "RFQ" model, where an existing agent requests a connection with another (really the same thing, just reversing the supplier/customer relationship).

Hmm. That could get tough to implement pretty quickly, eh? Designing an artificial economy sounds pretty unachievable (for the same reason that "designing" a real economy is unachievable), but hopefully the system could "self-assemble" if the agents could be designed simply enough and market-oriented enough.

Its come to my attention that Opencog (Goertzle's project) is using the "artificial economy" model for attention allocation; I'll have to read up on how they're approaching it.

Im also quite sure that there's been work done on agent based artificial economies outside of the AI feild, so I need to look into that too.

No comments: