Process and system including explainable prescriptions through surrogate-assisted evolution
OwnerHormoz Shahrzad, Babak Hodjatinventors
Life of this patent
Estimating the term…
Abstract
An explainable surrogate-assisted evolutionary optimization method, E-ESP, discovers rule-based decision strategies for which actions to take to achieve certain outcomes when historical training data is limited or unavailable. The resulting rules are human readable and thus facilitate explainability and trustworthiness unlike the black box solutions resulting from neural network solutions.
Claims11
A process for developing an optimized prescriptor model for determining optimal decision policy outcomes for a predetermined domain comprising:
- training a predictor surrogate model;
- and
- feeding the predictor surrogate model to an evolutionary algorithm framework for evolving a rule-based prescriptor model over multiple generations,
- wherein the rule-based prescriptor model has a general shape:
- coefficient1*inputi[lag1]p1{<|≤|>|≥}{coefficient2*inputj[lag2]p2|value}, where coefficients are numbers between zero and one, inputi and inputj are domain features, p1 and p2 are predetermined integer powers, and lags are values of the predetermined domain variables in previous time steps;
- wherein candidate rule-based prescriptor models are evaluated using the predictor surrogate model;
- and
- further wherein subsequent generations are evolved based on results from prior generations until an optimized rule-based prescriptor model is determined, wherein the optimized rule-based prescriptor model is selected based on a fitness evaluation.
A process for finding a decision policy that optimizes a set of outcomes to solve a known policy problem, comprising:
- i. training a predictor model based on known training data including context information (C), actions (A) performed in an included context, and outcome (O);
- ii. evolving a rule-based prescriptor model with the trained predictor model as a surrogate over a predetermined number of generations;
- iii. applying a determined elite evolved rule-based prescriptor model at a predetermined interval to a known policy problem to generate new decision policy data in the format (C, A, O);
- iv. collecting the new decision policy data (C, A, O) and adding the new decision policy data (C, A, O) to the known training data;
- v. repeating i. through iv. until a predetermined convergence metric is met;
- wherein the rule-based prescriptor model has a general shape:
- coefficient1*inputi[lag1]p1{<|≤|>|≥}{coefficient2*inputj[lag2]p2|value}, where coefficients are numbers between zero and one, inputi and inputj are domain features, p1 and p2 are predetermined integer powers, and lags are values of the predetermined domain variables in previous time steps;
- wherein an optimized rule-based prescriptor model is selected based on a fitness evaluation.
A method for storing instructions that, when executed by a computer, perform a process for developing an optimized prescriptor model for determining optimal decision policy outcomes for a predetermined domain comprising:
- training a predictor surrogate model;
- and
- feeding the predictor surrogate model to an evolutionary algorithm framework for evolving a rule-based prescriptor model over multiple generations,
- wherein the rule-based prescriptor model has a general shape:
- coefficient1*inputi[lag1]p1{<|≤|>|≥}{coefficient2*inputj[lag2]p2|value}, where coefficients are numbers between zero and one, inputi and inputj are domain features, p1 and p2 are predetermined integer powers, and lags are values of the predetermined domain variables in previous time steps;
- wherein candidate rule-based prescriptor models are evaluated using the predictor surrogate model;
- and
- further wherein subsequent generations are evolved based on results from prior generations until an optimized rule-based prescriptor model is determined, wherein the optimized rule-based prescriptor model is selected based on a fitness evaluation.
Dependencies are inferred from the claim text.