Multi-Task RL Policy Grouping to Reduce Parameter Interference

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing technologies fail to efficiently address the challenges of existing systems that use a single RL model for multiple tasks, leading to interference in shared parameters and negatively impacting task performance.

Innovation Solution

A computing device generates task-grouped RL policies by determining inter-task similarity scores and grouping tasks to minimize interference, using training circuitry to update RL models and form task groups with minimal negative impact on each other.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a single RL model with shared parameters is used to perform multiple tasks, then the system can handle diverse tasks with one model, but interference occurs in the shared parameters negatively impacting task performance

Engineering Contradiction:
Improvemulti-task capabilityVSAvoidtask performance
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent segments the single RL model into multiple task-specific RL models, each dedicated to performing a specific task. This segmentation eliminates parameter interference between tasks while maintaining the ability to handle diverse tasks through the collection of specialized models.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transitions from a single-dimension shared parameter space to a multi-dimension architecture where each task has its own parameter space. This dimensional separation allows tasks to be handled independently without interfering with each other's performance.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Reliability

If task grouping is implemented to reduce interference, then task performance improves, but the system complexity increases due to multiple RL models

Engineering Contradiction:
Improvetask performanceVSAvoidsystem architecture
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent creates a universal task management framework that can handle both simple and complex task scenarios. The system dynamically selects and coordinates task-specific models based on the current task requirements, providing a unified interface that masks the underlying complexity.

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

Solution Approach 2:

The patent implements dynamic task grouping and model selection mechanisms that adapt to different operational contexts. The system can dynamically adjust which task-specific models are active and how they are coordinated, optimizing performance while managing complexity through adaptive behavior rather than static rigid structures.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentUS20260003674A1Task grouping for reinforcement learning with multiple tasks
Publication Date: 2026.01.01 ATI TECHNOLOGIES ULC
  • US20260003674A1 patent drawing
  • US20260003674A1 patent drawing
  • US20260003674A1 patent drawing

AI summary

To generate reinforcement learning (RL) policies for the multiple tasks performable by a system, a computing device is configured to train an RL model for all tasks of a system to produce a general RL model. For each task, the computing device updates the parameters of the general RL model based on the task to produce a task-specific RL model. Based on comparisons of the general RL model to the task-specific RL models, the computing device determines inter-task similarity scores that represent the impact of a task on other tasks, the impact of other tasks on a task, or both. The computing device then groups the tasks of the system together based on the inter-task similarity scores and generates a task-grouped RL policy for each group of tasks.