Tuning Social Performance Index On Training-Data

To help a (perhaps artificial) decision-maker in compar- ing performance of its (perhaps artificial) social agents, a type of abstract evaluation-machine (structured on a Yaari-Quiggin functional- index) is presented. Designed on a certain sequence of increasingly severe goals O 0 , O 1 , ..., O l . . . ., O L , whenever applied to any agent A in a certain class D, it would evaluate its performance, taking into account conditions x := X, provided that value-gains parameters ω l|x := V al |x (O l ) − V al |x (O l−1 ) have been assigned, standardized on the behav- ior of a certain reference-agent A ∗ (chosen to represent a fair, wish- able behavior, by the designer point of view). The specific question of interest herein (frameable into machine learning) concerns a struc- tured, possibly automatizable, approach for extracting meaningful (nor- ∗ , to be assigned malized on assumed design-specifications) values ω |x to parameters ω l|x within a probabilistic setting, so that such values result properly tuned/calibrated on the training-data set χ[A ∗ ] associ- ated to the reference-agent A ∗ . In a Dirichlet-multinomial setup, a “prin- ciple of orthogonality” (applied to a certain updating variation operator Δ X [A ∗ ] T with respect to the sub-manifold associated to specifications of constraints on hyper-parameter space) is adopted to extract standard- ∗ which result tuned on training-data set χ[A ∗ ], structured ized values ω |xon a 3-way table which relates conditions to performance responses.