in the basic mathematical setup, they had an agent being a pair of a utility function and a strategy. The strategy doesn’t need to have anything at all whatsoever to do with the utility function.
Do you have a link to this usage of these terms? I think that, for rational agents, the strategy is an application of the utility function to the game and opponent(s). They’re not quite isomorphic—you can derive the strategy from the utility function, but not the reverse (you can derive a utility theory, but it may be underspecified and not the actual agent’s utility theory outside this game). Similary, Oracles don’t add anything to the theory—they’re just an implementation of higher-levels of simulation/prediction.
I think a lot of the recursion falls apart when you add in resource constraints, via a cost to calculation. When the rewards are reduced by the amount of energy taken for the computation, it’s a lot harder to unilaterally optimize, and it becomes worth it to accept lower but “safer” payouts.
Do you have a link to this usage of these terms? I think that, for rational agents, the strategy is an application of the utility function to the game and opponent(s). They’re not quite isomorphic—you can derive the strategy from the utility function, but not the reverse (you can derive a utility theory, but it may be underspecified and not the actual agent’s utility theory outside this game). Similary, Oracles don’t add anything to the theory—they’re just an implementation of higher-levels of simulation/prediction.
I think a lot of the recursion falls apart when you add in resource constraints, via a cost to calculation. When the rewards are reduced by the amount of energy taken for the computation, it’s a lot harder to unilaterally optimize, and it becomes worth it to accept lower but “safer” payouts.