Reward Estimation Model Using Privileged Training Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing computer models face challenges in making accurate predictions due to limited access to information at the time of prediction and delayed feedback, particularly in reinforcement learning systems, leading to inaccurate efficiency estimations.

Innovation Solution

A combined machine learning model incorporating learning using privileged information (LUPI) and multi-armed bandits (MAB) paradigms to enhance reward estimation by leveraging internal data, generating estimated replacements for delayed metrics.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If reinforcement learning systems wait for delayed reward feedback to make predictions, then prediction accuracy may improve, but response time becomes unacceptable

Engineering Contradiction:
Improveprediction accuracyVSAvoidresponse time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary actions by training a predictive model during the training phase to estimate future rewards before actual feedback is received. This allows the RL system to make informed decisions immediately without waiting for delayed reward signals, thus maintaining fast response time while improving prediction accuracy through the pre-trained model's estimates.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

A predictive model serves as an intermediary between the RL system and the delayed reward feedback. This intermediary model estimates the future reward values based on current state and action, allowing the RL system to operate with immediate feedback approximations rather than waiting for actual delayed rewards, thereby resolving the time-accuracy tradeoff.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If expert-built systems use crafted rules for decision-making, then transparency is improved, but flexibility and adaptability to available data are reduced

Engineering Contradiction:
ImprovetransparencyVSAvoidflexibility
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The system merges the transparency advantage of expert-built rule systems with the adaptability of machine learning models. The RL framework maintains interpretable decision structures while incorporating trained predictive models that can adapt to data patterns, achieving both transparency through the decision framework and flexibility through the learned predictive capabilities.

Inventive Principle:
Principle #5Merging (Combining)

3Ease of operation

If reinforcement learning systems operate without privileged information, then ease of operation is improved, but measurement precision of reward estimation deteriorates

Engineering Contradiction:
Improveease of operationVSAvoidreward estimation accuracy
Core Design Contradiction:
Ease of operationVSMeasurement precision

Solution Approach 1:

Privileged information is utilized during the preliminary training phase to teach the predictive model accurate reward estimation. Once trained, the model operates independently during execution without requiring access to privileged information, thus maintaining ease of operation while achieving high reward estimation accuracy through the pre-acquired knowledge.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4657339A1Training resource allocation models
Publication Date: 2025.12.03 INTUIT INC
  • EP4657339A1 patent drawingFigure 1A
  • EP4657339A1 patent drawingFigure 1B
  • EP4657339A1 patent drawingFigure 2

AI summary

A method includes applying a training controller to an untrained predictive model and training data to generate a trained predictive model. The training data includes privileged information. Applying the training controller is an iterative process that repeats until convergence and includes a loss function determination phase, an update phase that updates the untrained predictive model, and a test phase that includes applying the training data to an updated version of the untrained predictive model. The privileged information is applied during the loss function determination phase and excluded during the test phase. The method also includes integrating the trained predictive model into a reward estimation function to generate a trained combined model. The trained predictive model's output includes an input to the reward estimation function. The trained predictive model's output includes a reward determination for the reward estimation function. The trained combined model is presented.