Marginal Ratio Estimator for Contextual Bandit Policy Evaluation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing off-policy evaluation techniques for contextual bandit systems suffer from high variance and inefficiency, particularly when there are substantial policy shifts, making it challenging to compare and optimize decision-making policies.

Innovation Solution

The Marginal Ratio (MR) estimator is introduced, which calculates a policy ratio and uses machine learning to train a parametric model that estimates the expectation of outcomes for a to-be-evaluated policy, reducing variance and improving processing efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If existing off-policy evaluation techniques are used, then policy evaluation can be performed, but the variance is high and processing efficiency is poor

Engineering Contradiction:
Improveevaluation accuracyVSAvoidprocessing efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent introduces an intermediate estimator that serves as a mediator between the importance sampling estimator and the outcome model. This intermediate estimator decomposes the evaluation into manageable components, reducing variance while maintaining computational efficiency. The intermediate estimator acts as a bridge that allows accurate policy evaluation without requiring full retraining of outcome models for each policy comparison.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent segments the policy evaluation process into distinct components: importance sampling weights, intermediate estimators, and outcome model predictions. By dividing the evaluation into these separate segments, each component can be optimized independently, reducing overall variance and improving processing efficiency compared to monolithic evaluation approaches.

Inventive Principle:
Principle #1Segmentation

2Adaptability or versatility

If existing off-policy evaluation techniques are used, then policy comparison is possible, but the technique scales poorly with increasing policy shifts

Engineering Contradiction:
Improvepolicy comparison capabilityVSAvoidcomputational complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent performs preliminary actions by pre-computing importance sampling weights and training outcome models on the reference policy before actual policy evaluation. This preliminary preparation allows for efficient comparison of multiple policies against the reference policy without repeating computationally intensive operations, enabling better scaling with policy shifts.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent changes parameters by using importance sampling weights that adapt to different policy distributions. Instead of retraining models for each policy, the method adjusts the evaluation using parameter changes in the importance weights, allowing efficient comparison across diverse policies with varying degrees of shift from the reference policy.

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If dataset reCollection is performed for each policy comparison, then accurate evaluation is achieved, but time and resources are significantly consumed

Engineering Contradiction:
Improveevaluation accuracyVSAvoiddata collection time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent implements feedback by using the reference policy dataset multiple times for evaluating different policies. The same dataset provides feedback for comparing multiple policies against the reference, eliminating the need to collect new data for each policy while maintaining evaluation accuracy through importance sampling and intermediate estimators.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent creates virtual copies of the reference policy dataset through importance sampling weights, allowing the same physical dataset to serve multiple evaluation purposes. Instead of collecting separate datasets for each policy comparison, the method creates weighted copies of the reference data that reflect different policy distributions, saving time and resources.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4557180A1Method of policy evaluation for contextual bandits
Publication Date: 2025.05.21 LEMON INC(GB)
  • EP4557180A1 patent drawingFigure 1a
  • EP4557180A1 patent drawingFigure 1b
  • EP4557180A1 patent drawingFigure 2a~2b

AI summary

A computer-implemented method of policy evaluation for a contextual decision-making policy, wherein the contextual decision-making policy is used to make one or more decisions, where the outcome of each decision depends upon a context in which the decision is made, the method comprising: obtaining an existing contextual decision-making policy; obtaining a first dataset of decisions made using the existing contextual decision-making context, wherein the first dataset further includes the respective outcome and context of each decision; obtaining a to-be-evaluated contextual decision-making policy; using machine learning to train an outcome model for the to-be-evaluated contextual decision-making policy, based on the existing contextual decision-making policy, the to-be-evaluated contextual decision-making policy and the first dataset; and estimating an expectation for the to-be-evaluated contextual decision-making policy, based on the outcome model.