Thread by YK Community Feed
- Tweet
- May 7, 2022
- #GameTheory #MachineLearning
Thread
As Machine Learning predictions become more accurate, the assumptions behind Nash game theory are pushed to the limits of their validity because of the prediction<->decision feedback loop. Game theory needs to be adapted with a shift of paradigm.
In Nash game theory, it is assumed that the agents' decisions are counterfactually independent from the past and the present. The Nash equilibrium is defined as stable given unilateral deviations.
But in a world in which each click is predicted, this assumption needs to be lifted. One way of doing so is to assume instead that predictions are fully counterfactually dependent on the decisions, which is known as Perfect Prediction.
In addition, Common Knowledge of Rationality (I know they know I know... they are rational) is extended to the more stringent assumption of Necessary Rationality: we are all rational in all possible worlds, even if our decisions had been different.
New equilibria can be defined that are not subsets of Nash but instead generalize superrationality. They are at most unique and Pareto-optimal. A first analysis shows that social utility metrics are higher than those of Nash for the studied games.