Meta-Gradient Return Function Training for Faster Reinforcement Learning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing reinforcement learning systems face inefficiencies in training, particularly in converging on optimal policy parameters, leading to suboptimal performance and requiring extensive data for effective learning.

Innovation Solution

Applying meta-learning with a meta-objective function to iteratively update both policy and return parameters, using meta-gradients to adapt the return function during training, allowing for faster convergence on optimal policy parameters.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If traditional reinforcement learning training methods are used, then the system can learn from experience, but the convergence on optimal policy parameters is slow and requires extensive data

Engineering Contradiction:
Improvetraining efficiencyVSAvoidconvergence time
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The patent applies preliminary action by using meta-gradients to pre-adjust return parameters before full training convergence. The meta-objective function computes gradients of the policy objective with respect to return parameters, allowing the system to proactively optimize return function parameters based on anticipated training needs, thereby accelerating convergence without requiring extensive training data

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements feedback by using the meta-objective function to continuously evaluate how changes in return parameters affect policy performance. The meta-gradients provide feedback signals that guide the optimization of return parameters, creating a closed-loop system where the policy training process itself informs the optimization of the return function, leading to faster convergence

Inventive Principle:
Principle #23Feedback

2Measurement precision

If traditional reinforcement learning training methods are used, then the system can update policy parameters, but the learning process is inefficient and requires extensive data

Engineering Contradiction:
Improvelearning accuracyVSAvoiddata requirements
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent applies parameter changes by optimizing return parameters (such as discount factor γ and bootstrapping parameter λ) alongside policy parameters. The meta-gradient method computes ∂J/∂η where J is the policy objective and η are return parameters, enabling the system to learn optimal parameter configurations that improve learning accuracy while reducing the amount of training data needed

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The system performs preliminary optimization of return parameters using meta-gradients before full policy convergence is achieved. This preliminary adjustment of return function parameters based on meta-objective feedback allows the system to reach higher learning accuracy with fewer training samples by ensuring the return function is well-tuned early in the training process

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3769264B1Meta-gradient updates for training return functions for reinforcement learning systems
Publication Date: 2025.08.27 GDM HOLDING LLC
  • EP3769264B1 patent drawingFigure 1
  • EP3769264B1 patent drawingFigure 2
  • EP3769264B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for reinforcement learning. The embodiments described herein apply meta-learning (and in particular, meta-gradient reinforcement learning) to learn an optimum return function G so that the training of the system is improved. This provides a more effective and efficient means of training a reinforcement learning system as the system is able to converge on an optimum set of one or more policy parameters θ more quickly by training the return function G as it goes. In particular, the return function G is made dependent on the one or more policy parameters θ and a meta-objective function J' is used that is differentiated with respect to the one or more return parameters η to improve the training of the return function G.