Prediction Program Policy Feasibility Validation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing prediction methods using machine learning models often present policies that are unlikely to be implemented, leading to the omission of feasible policies and the presentation of impractical ones, as they focus on action-value functions rather than practical implementation feasibility.
Innovation Solution
A prediction program and apparatus that lists correlated feature amounts, creates policies based on differences between these amounts, and determines policy appropriateness using past performance information, treating uncontrollable variables appropriately to ensure feasibility.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a reinforcement learning method is applied to predict policies by maximizing action-value functions, then prediction accuracy for optimal policies is improved, but the feasibility of presented policies deteriorates as impractical policies are selected
Solution Approach 1:
The patent introduces performance information as an intermediary factor that mediates between the action-value function maximization and policy feasibility. The determination unit uses performance information (past performance data) to filter and validate policies, ensuring that only feasible policies are presented while maintaining high prediction accuracy through the combined use of action-value functions and performance-based validation.
2Productivity
If only action-value functions are used to determine policies, then computational efficiency is improved, but the consideration of uncontrollable variables deteriorates leading to impractical policies
Solution Approach 1:
The patent segments the policy determination process into two distinct stages: (1) a search process using action-value functions to generate candidate policies, and (2) a determination process using performance information to filter and validate policies. This segmentation allows the system to maintain computational efficiency in the search phase while thoroughly considering controllable and uncontrollable variables in the determination phase.
Solution Approach 2:
The patent implements feedback by using performance information (past performance data) to validate and filter generated policies. The determination unit receives performance feedback and adjusts the policy selection accordingly, ensuring that policies are not only computationally efficient but also practically feasible by accounting for uncontrollable variables that have historically impacted performance.
Data Source
AI summary
A non-transitory computer-readable storage medium storing a prediction program for causing a computer to perform processing including: listing combinations of feature amounts that are correlated with a target label; creating a policy to achieve the target label for a prediction target based on a difference between the listed combinations of the feature amounts and a combination of feature amounts of the prediction target; and determining appropriateness of the created policy based on performance information that indicates past performances.