Aside from the issue that dollars ought to have diminishing returns in utility of some form, expected utility is perfectly capable of modeling situations where ergodicity/ensemble assumptions are violated and a greedy one-step policy of making the decision with maximum immediate expected utility doesn’t work. You just need to actually include the full dynamics of the situation in your model and plan appropriately.
I realize that this may be shocking, but in order to get a correct answer, you need to ask the correct question. ‘Pray, Mr Babbage, if you put the wrong numbers in, will the right numbers come out?’ You may want to study the uses of expected utility or decision theory a little bit more before trying to refute and replace it...
For example, if you have a limited amount of money and are vulnerable to gambler’s ruin, expected utility will still give you the correct answer if you set up a environment model and do planning or backward induction to calculate the optimal policy (which converges on Kelly criterion with longer horizons, and Kelly converges on greedy 1-step EU-maximization with larger capital). I’ve analyzed one game where greedy 1-step EU-maximization almost always leads to zero gains but a decision tree—using nothing but expected utility maximization and a correct model of the game—leads to maximal gains >94% of the time: https://www.gwern.net/Coin-flip Nothing new here, backwards induction and expected utility go back at least to von Neumann/Morgenstern.
Aside from the issue that dollars ought to have diminishing returns in utility of some form, expected utility is perfectly capable of modeling situations where ergodicity/ensemble assumptions are violated and a greedy one-step policy of making the decision with maximum immediate expected utility doesn’t work. You just need to actually include the full dynamics of the situation in your model and plan appropriately.
I realize that this may be shocking, but in order to get a correct answer, you need to ask the correct question. ‘Pray, Mr Babbage, if you put the wrong numbers in, will the right numbers come out?’ You may want to study the uses of expected utility or decision theory a little bit more before trying to refute and replace it...
For example, if you have a limited amount of money and are vulnerable to gambler’s ruin, expected utility will still give you the correct answer if you set up a environment model and do planning or backward induction to calculate the optimal policy (which converges on Kelly criterion with longer horizons, and Kelly converges on greedy 1-step EU-maximization with larger capital). I’ve analyzed one game where greedy 1-step EU-maximization almost always leads to zero gains but a decision tree—using nothing but expected utility maximization and a correct model of the game—leads to maximal gains >94% of the time: https://www.gwern.net/Coin-flip Nothing new here, backwards induction and expected utility go back at least to von Neumann/Morgenstern.
The article was admittedly premature.