Meta-Gradient Options for Stable Multi-Task Reinforcement Learning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing reinforcement learning systems struggle to efficiently learn sequences of actions that are useful across multiple tasks, often requiring hand-engineered solutions and lacking data efficiency and stability, especially in real-world applications.

Innovation Solution

A reinforcement learning system that learns options through meta-gradients, utilizing a manager neural network and a set of option policy neural networks to discover sequences of primitive actions that are useful across different tasks, with training mechanisms that include model-based and model-free methods like actor-critic and Deep Q-Network techniques.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If reinforcement learning systems use hand-engineered solutions for learning action sequences, then the system can perform multiple tasks, but the system lacks data efficiency and stability

Engineering Contradiction:
Improvemulti-task capabilityVSAvoidlearning stability
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The system segments the learning process into two distinct levels: option-level learning (learning sequences of primitive actions) and task-level learning (learning which options to apply to which tasks). This segmentation allows each level to specialize and stabilize independently, with option policies being reused across multiple tasks, thereby improving overall learning stability and data efficiency while maintaining multi-task versatility.

Inventive Principle:
Principle #1Segmentation

2Manufacturing precision

If reinforcement learning systems learn task-specific policies separately, then each task can be optimized, but computational resources and data requirements increase

Engineering Contradiction:
Improvetask optimizationVSAvoidcomputational resources
Core Design Contradiction:
Manufacturing precisionVSUse of energy by moving object

Solution Approach 1:

The system learns a set of universal option policies that can be applied across multiple tasks. Each option policy represents a reusable sequence of primitive actions that can benefit multiple tasks. This universality reduces computational resources and data requirements because the system doesn't need to learn the same action sequences repeatedly for each task - instead, it learns once and reuses across tasks while still achieving task-specific optimization through the task-level policy.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Adaptability or versatility

If reinforcement learning systems learn complex action sequences, then the system can perform sophisticated tasks, but the learning process becomes inefficient and resource-intensive

Engineering Contradiction:
Improvetask complexityVSAvoidlearning efficiency
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The system performs preliminary learning of option policies at the option level before applying them to specific tasks. By pre-learning reusable action sequences and their associated rewards at the option level, the system eliminates the need to relearn these sequences for each task. This preliminary action significantly improves learning efficiency when dealing with complex tasks, as the system can compose sophisticated behaviors from pre-learned options rather than learning everything from scratch.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4111383B1Learning options for action selection with meta-gradients in multi-task reinforcement learning
Publication Date: 2025.08.06 GDM HOLDING LLC
  • EP4111383B1 patent drawingFigure 1
  • EP4111383B1 patent drawingFigure 2
  • EP4111383B1 patent drawingFigure 3

AI summary

A reinforcement learning system, method, and computer program code for controlling an agent to perform a plurality of tasks while interacting with an environment. The system learns options, where an option comprises a sequence of primitive actions performed by the agent under control of an option policy neural network. In implementations the system discovers options which are useful for multiple different tasks by meta-learning rewards for training the option policy neural network whilst the agent is interacting with the environment.