Hayek

I’ve been reading What is Thought? by Eric Baum, a book on the mind by an AI researcher.

For part of the book, Baum describes solving Blocks World problems with a computational economics model. The problem is called Hayek, and it is described in this article. I will just describe the part that is most interesting to me.


The program is implemented as a collection of agents which bid for ownership of the puzzle. Only the owner of the puzzle at any given time is allowed to move its pieces. The puzzle itself is turned into a cash reward when it is solved. Otherwise the world is zero-sum between agents, with money only changing hands during the sales of the puzzle between agents.

Using such rules, evolutionary algorithms are applied to encourage the development of more profitable agents. Successful agents were those able to buy low and sell high, which (assuming accurate pricing) would mean being able to advance the solution the most.

Of course, accurate pricing is a hairy issue; the hardcoding of a good scoring function helps the system a lot.

This is a general parallelization paradigm, where heuristics exist in ecosystems which estimate the value of each other’s actions and compete to maximize their own value in that context. An agent’s pricing of the puzzle state is subjective to the extent that it is specific to the position of that agent’s actions in the chain leading to the solution, and objective to the extent that the other agents see the same chain. I would guess that it is more objective when the action chain is serial and more subjective when many alternative parallel chains exist.

I find Economic AI models cool because they clearly work for humans, and if could be made to work for computers would allow the construction of hybrid cybernetic economies where humans and agents can compete in the same pool.

Comments are closed.