So how does this work?
Consider the architecture of a typical processing scheme that maps inputs onto certain values
derived from processed outputs as conditioned by the parameters of some predictive model:
latter are the sort of parameters that conventional neural networks exploit.
are instead “guided” by an oracle � and its dual ∗� respectively:
those otherwise inaccessible states. A possible demonstration that this oracle exists “in principle” can be glimpsed by equating it with an optimization over all future states of the game. This approach relies on transferring the inaccessibility of other private states to the future, where those once inaccessible states become revealed through play over time … in the future.
Note that playing into the future is exactly what deep learning model training purports to do; but for sufficiently complex negotiations, this training simply cannot provide a sufficiently dense subset of possible plays. Consequently, no conventional machine learning system can qualify as an oracle.
Alternatively, imagine swallowing this essential inaccessibility in hyperbolic 3-space. There’s a lot of room in hyperbolic space…. Anyway, that’s the sort of thing that an oracle might do.
It’s what IRIEã actually does.