Fuzzy-Inference Task Allocation for Multi-Agent Sub-Task Conflicts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing multi-agent reinforcement learning methods struggle to quickly and accurately allocate functionally-adaptive sub-tasks to agents in complex cooperative tasks due to increased complexity, often leading to inefficient task execution and competitive sub-task execution among agents.
Innovation Solution
A multi-agent task allocation method based on fuzzy inference, utilizing Gaussian fitting for adaptive decomposition, dual-time-scale layered cooperative architecture, and temporal difference loss functions to optimize sub-task and agent policy networks, ensuring accurate and efficient task allocation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If existing multi-agent reinforcement learning methods are used for task allocation in complex cooperative tasks, then the system can handle basic cooperative tasks, but the task allocation efficiency and accuracy deteriorate due to increased complexity
Solution Approach 1:
The patent divides the complex task allocation problem into two distinct time scales: a fast time scale for real-time sub-task assignment and a slow time scale for policy optimization. This segmentation allows the system to handle complex tasks efficiently by separating immediate allocation decisions from longer-term learning processes, thereby maintaining high productivity while dealing with increased task complexity.
Solution Approach 2:
The patent implements dynamic task allocation by continuously updating sub-task assignments based on current agent states and task requirements. The fast time scale enables real-time adaptation to changing conditions, while the slow time scale gradually optimizes policies. This dynamic approach allows the system to maintain high adaptability and efficiency even as task complexity increases.
2Ease of operation
If bottom-to-top task selection processes are used, then agents can select sub-tasks based on local information, but task allocation accuracy deteriorates due to lack of top-to-bottom optimization
Solution Approach 1:
The patent implements a feedback mechanism where the slow time scale policy optimization provides top-to-bottom guidance to the fast time scale sub-task selection. The policy network learns from accumulated experience and provides improved allocation strategies that enhance accuracy. This feedback loop maintains the simplicity of local sub-task selection while progressively improving overall allocation accuracy through learned policies.
3Extent of automation
If agents independently select sub-tasks based on local information, then the system operates with decentralized control, but task allocation accuracy deteriorates due to competitive execution of same sub-tasks
Solution Approach 1:
The patent merges decentralized sub-task selection with centralized policy optimization by combining the fast autonomous agent decisions with the slow learned policy guidance. The policy network, trained through centralized training, coordinates agent behaviors to avoid competitive execution of the same sub-tasks while maintaining decentralized operation during execution. This merging approach preserves autonomous operation while improving allocation accuracy.
Data Source
AI summary
Provide is a multi-agent task allocation method based on fuzzy inference, which includes: obtaining multidimensional features of all agents at a historical moment and multidimensional features of all sub-tasks at the historical moment, and determining a sub-task selector network based on fuzzy inference according to mean values and covariances of all sub-tasks; with a goal of minimizing a first Temporal Difference (TD) loss function, training the sub-task selector network based on fuzzy inference and a sub-task policy network by utilizing a sub-task evaluation network; with a goal of minimizing a second TD loss function, training an agent policy network by utilizing an agent credit allocation network; sequentially inputting locally-observed information of each agent at a current moment, an execution action at a previous moment, and a sub-task at the previous moment into a trained agent policy network and a trained sub-task selector network based on fuzzy inference for sub-task allocation.