Chatty Twins

Project status

Ongoing

Introduction

Chatty Twins aims to test the feasibility of deploying end-to-end Supply Chain Digital Twins (SC-DT), by building a unique set of re-usable multi-agent capabilities that they need to encapsulate and test them with use cases inspired by industry.
To do so, the project brings together academics and industry experts to research at the intersection of digital twins, LLMs, multi-agent AI systems and supply chains.
 

Explaining the science

Digital Twins are digital models that are constantly updated with data to emulate as well as simulate real objects and processes, making them a valuable information source for decision-making. Combining DT with AI can enable decision-making at speed and scale beyond human capability. DT have been popular in manufacturing but the commercial interdependent nature of SC means that data within them is siloed and inaccessible by SC members. This is especially relevant for distributed decision-making spanning multiple companies where AI systems could represent companies and interact with each other to achieve improved outcomes for all participants.
 

Recent advances in LLMs have shown their potential to perform industrial control tasks with far fewer samples and easier implementation than extant approaches, making them more accessible to SMEs. LLM frameworks can use dedicated tools for specialized tasks (e.g., a sales forecasting model) and perform multi-agent LLM interactions to achieve goals together. While initial results of multi-agent LLM frameworks are promising, challenges remain to enable distributed decision-making in practice. In this project we address the key challenges of AI agents determining dependencies in partially observable supply chains, efficient methods to keep the tools LLMs rely on up to date in the dynamic environment of supply chains using machine unlearning, and leveraging LLMs to arrive at decisions that benefit the overall supply chain.
 

Project aims

  1. Create and validate a framework to automatically determine dependencies between SC-DT.
  2. Establish and validate a suite of methods to enable SC-DT to arrive at near-optimal solutions in problems that necessitate distributed decision-making on the pathway of dependencies.
  3. Empirically validate the framework and assess its generalisability across a range of distributed supply chain decision-making scenarios.

Applications

The outcome of the project will benefit:

  1. Industry, by providing an overview of the current and emerging research landscape in AI for supply chain management, developing use cases and potential areas of opportunity and risk.
  2. The academic community, by exploring the feasibility and challenges associated with the use of LLMs for distributed decision making in practical contexts.
     

Recent updates

Fast Machine Unlearning Without Retraining Through Selective Synaptic Dampening

 

Identifying contributors to supply chain outcomes in a multi-echelon setting: a decentralised approach

Organisers

Collaborators

Previous contributors

Funders