MARL Policy Model for Inter-Agent Communication Control

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing multi-agent reinforcement learning (MARL) systems face challenges in efficiently managing inter-agent communication due to resource constraints such as bandwidth and power capacity, leading to costly unconstrained communication in applications like the Internet of Things and robotics.

Innovation Solution

A MARL system that trains agents to control when and with whom to communicate by using aggregated reward signals combining task rewards and communication penalties, optimizing communication through a policy model that determines whether and to whom to send messages based on a reward score that balances communication value and cost.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If unconstrained inter-agent communication is implemented, then agents can share information and agree on joint strategies, but resource consumption (bandwidth, power capacity) increases significantly

Engineering Contradiction:
Improveinformation sharing effectivenessVSAvoidcommunication resource consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent implements dynamic communication control where agents adapt their communication behavior based on real-time task requirements and resource conditions. The policy model dynamically determines communication actions (send, receive, or stay silent) rather than using fixed communication patterns, allowing the system to optimize the balance between information sharing and resource consumption during task execution

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent changes the communication parameter from unconstrained to constrained by introducing a policy model that regulates communication based on task relevance. The reward function modifies communication behavior by penalizing unnecessary communications and rewarding task-effective information exchange, effectively changing the communication parameter from binary (communicate/don't communicate) to a continuous optimization problem

Inventive Principle:
Principle #35Parameter changes

2Loss of information

If agents communicate with all other agents at every step, then complete information exchange is achieved, but communication cost increases

Engineering Contradiction:
Improveinformation completenessVSAvoidcommunication cost
Core Design Contradiction:
Loss of informationVSLoss of energy

Solution Approach 1:

The patent extracts only the necessary communication interactions from the set of all possible agent interactions. The policy model evaluates each potential communication action and selectively executes only those that are task-relevant, removing unnecessary communications from the system while preserving essential information exchange pathways

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies partial action by having agents communicate only when the expected information gain exceeds a threshold, rather than communicating with all agents at all times. The policy model determines the optimal subset of communication actions needed to achieve task objectives without the excess cost of complete communication

Inventive Principle:
Principle #16Partial or excessive action

3Use of energy by moving object

If communication is minimized to save resources, then resource efficiency improves, but task coordination capability deteriorates

Engineering Contradiction:
Improveresource efficiencyVSAvoidtask coordination capability
Core Design Contradiction:
Use of energy by moving objectVSReliability

Solution Approach 1:

The patent implements feedback through the reward function that evaluates communication actions based on their impact on task progress. The reward signal provides feedback to the policy model, reinforcing communication behaviors that improve task coordination while penalizing communications that do not contribute to task objectives, enabling the system to learn optimal communication strategies that balance resource efficiency and coordination capability

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent enables agents to self-regulate their communication behavior through the trained policy model. Each agent independently determines its own communication actions based on its local state and learned policies, allowing the system to achieve efficient coordination without centralized control while maintaining resource efficiency

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4195121A1System and method for controlling inter-agent communication in multi-agent systems
Publication Date: 2023.06.14 INTEL CORP
  • EP4195121A1 patent drawingFigure 1
  • EP4195121A1 patent drawingFigure 2
  • EP4195121A1 patent drawingFigure 3

AI summary

An agent in a multi-agent system is provided with a policy model that controls communication of the agent with other agents in the multi-agent system. The policy model is trained by using MARL. The policy model receives more messages from one or more other agents in the multi-agent system. The policy model generates a reward score based at least on a hidden state of the agent and the one or more messages. The reward score represents an aggregation of a value of sending the message for a task and a cost of sending the message. The policy model determines whether to send the message based on the reward score. After determining to send the message, the policy model generates the message based on the hidden state of the agent and the one or more messages and sends the message to one or more other agents in the multi-agent system.