Relational Forward Model for Multi-Agent Prediction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current neural network systems struggle to effectively predict and explain the actions of multiple agents in shared environments, particularly in reinforcement learning scenarios, as they require extensive data and computational resources and often fail to account for the influences between agents accurately.

Innovation Solution

A relational forward model (RFM) neural network system that processes agent data using an encoder graph neural network to generate encoded graph data, which is then processed by a recurrent graph neural network to update node and edge attributes, providing representation data that predicts agent actions and influences, facilitating improved performance and cooperation among agents.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional neural network systems are used to predict actions of multiple agents, then prediction capability is achieved, but data requirements and computational resources increase extensively

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

Solution Approach 1:

The system segments the multi-agent environment into a graph structure where individual agents and environmental entities are represented as nodes, and their relationships as edges. This segmentation allows the neural network to process information in structured units rather than as a monolithic input, reducing the overall data processing burden while maintaining prediction accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transforms the traditional flat input data structure into a graph-based representation with nodes and edges, adding a structural dimension to the data. This dimensional transformation enables the network to capture relational information more efficiently, reducing the quantity of data needed to represent complex multi-agent interactions.

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

2Measurement precision

If traditional neural network systems are used to predict actions of multiple agents, then prediction capability is achieved, but computational resources increase extensively

Engineering Contradiction:
Improveprediction accuracyVSAvoidcomputational resources
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

By segmenting the problem into graph-based relational representations, the system enables more efficient computational processing. The graph structure allows for localized computations at nodes and edges rather than global matrix operations, significantly reducing computational resource requirements while preserving prediction accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes the parameter representation from traditional dense vectors to graph-based structural parameters (nodes, edges, and their attributes). This parameter transformation enables the neural network to process multi-agent interactions with fewer computational operations, reducing energy and resource consumption.

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If traditional neural network systems are used, then basic prediction is achieved, but influences between agents are not accounted for accurately

Engineering Contradiction:
Improveprediction accuracyVSAvoidinter-agent influence information
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent adds a relational dimension by representing inter-agent influences as edges in the graph structure. This dimensional addition explicitly models the influence relationships between agents, preventing information loss about how agents affect each other's actions while improving prediction accuracy.

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

Solution Approach 2:

The system segments inter-agent influences into discrete edge relationships in the graph, allowing each influence to be independently represented and processed. This segmentation enables the network to accurately capture and process influence information without losing it in aggregated traditional input representations.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3776370B1Graph neural network systems for behavior prediction and reinforcement learning in multiple agent environments
Publication Date: 2024.11.20 DEEPMIND TECH LTD
  • EP3776370B1 patent drawingFigure 1
  • EP3776370B1 patent drawingFigure 2a
  • EP3776370B1 patent drawingFigure 2b

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for predicting the actions of, or influences on, agents in environments with multiple agents, in particular for reinforcement learning. In one aspect, a relational forward model (RFM) system receives agent data representing agent actions for each of multiple agents and implements: an encoder graph neural network subsystem to process the agent data as graph data to provide encoded graph data, a recurrent graph neural network subsystem to process the encoded graph data to provide processed graph data, a decoder graph neural network subsystem to decode the processed graph data to provide decoded graph data and an output to provide representation data for node and/or edge attributes of the decoded graph data relating to a predicted action of one or more of the agents. A reinforcement learning system includes the RFM system.