Intelligent agents take control

Tom Shelley reports on a radical approach to manufacturing and management where each order, part and component is responsibly intelligent.

In systems being trialled and developed, every order, product, assembly, and machine is associated with its own intelligent agent, able to communicate and negotiate with others to obtain best solutions.

Products can find the most appropriate machines to make them, and reach them by the best route, quickly adapting to the break down of trucks or conveyors, or new machines suddenly coming online.

First real world applications all seem to be in defence, but commercial blue chip commercial customers are close behind, looking for ways to avoid the logjams and software configuration delays that bedevil conventional supply chain and manufacturing management processes.

The idea of components and machines being associated with intelligent 'agents' - some people call them 'holons' - has been researched for some time. The ancestor of all agents is named Eliza, born at MIT in 1966 thanks to Professor Joseph Weizenbaum. Written in only a couple of hundred lines of code, she was created in order to simulate a conversation between a patient and a Rogerian psychotherapist. In 1977, Carl Hewitt, also at MIT, proposed the concept of an agent with autonomous behaviour, able to answer requests from other agents. Much of the subsequent development has been driven by the US and UK military, seeking ways to ensure that swarms of unmanned aerial vehicles achieve their objectives, even when some are destroyed or damaged, and to improve the realism of simulated battle and the management of real battles.

In the commercial field, a component or whole car that is to be painted blue can be associated with a piece of software that knows this. The software agent will negotiate with conveyors and paint shops to send the item to a paint shop that can do blue, by the shortest and most economic route. Should a conveyor break down, or the paint shop run out of blue paint, the component can negotiate an arrangement with a new paint shop and alternative conveyors, but will inform all the other components about the problem, according to a subscriber list of agents that 'need to know'.

Current state of the art has just been revealed at a workshop in Cambridge. "Intelligent Agents in Industrial Control" was held at the Centre for Distributed Automation and Control (CDAC), part of the University of Cambridge Institute for Manufacturing. Dr Duncan McFarlane, director of the Centre, presided over the proceedings and explained some of the advantages of an agent approach.

He considered that the present hierarchical approach to manufacturing and production is too restrictive in the event of disruptions. If an individual machine goes out of service, or a new one is installed in a conventional environment, modifications have to be made to software in planning, scheduling and manufacturing order layers. With an agent based approach, in what he calls, a Holonic Manufacturing System, new pieces of equipment just have to be installed in a 'Plug and Play' manner, after which the whole system will find them and make use of them. He proposes that a software agent be assigned to each order, machine (resource) and physical product. The agent and the physical item it is associated with is called a 'holon' - an autonomous but integratable entity. The order holon negotiates with resource holons to generate a product holon, which comprises the product itself, information about the product required and the software agent responsible for getting the product made. In the event of broken resource, it renegotiates with the other resources to still get itself made.

Professor Vladimir Marik, managing director of the Rockwell Automation Research Center in Prague, told Eureka that Europe is well ahead of the US in implementing agents in commercial supply chains and manufacturing. Much of the best work, apparently, is presently being undertaken in the UK.

Professor Marik described simulated manufacturing systems based on agents written in JADE, Java Agent DEvelopment framework, developed by Telecom Italia Labs. However, he also happened to mention that Rockwell had a demonstration agent based system running in the US, to improve the fault tolerance of weapons cooling systems on one of the US Navy's aircraft carriers. Written with C++ as its core language, it comprises 116 agents running on five Rockwell ControlLogix and one FlexLogix controllers. While the C++ software evidently works, moving to Java will allow the software to run in less memory space and make use of established libraries and web enabled links to ERP systems. Java is also unaffected by operating systems, or operating system versioning. Professor Marik told Eureka that commercial customers were presently "considering" taking up the Rockwell version of the technology but declined to give details.

In the manufacturing arena, software obviously cannot run on individual components being made, so it has to be closely associated with them. This brings together the development of agents, and the accurate recognition of individual components at all stages as they progress through a manufacturing process and/or supply chain.

Based within CDAC is the European research arm of the Auto-ID Center, founded in 1999. The Center is a partnership between more than 50 global companies plus the University of Cambridge, MIT, the University of Adelaide, St Gallen University in Switzerland, Keio University in Japan and Fudan University in China. Its aim is to develop a standardised technology based on very low cost RFID, Radio Frequency IDentification. RFID was first conceived by Harry Stockman in the US in 1948, but it was not until the 1970s that technology developed to the point that it became commercially feasible. The Auto ID Center has aggressively driven down the price of RFID tags. Antennae are now generally printed on flexible circuit boards, and developments by companies such as Plastic Logic, which include the ink jet printing of associated electronic circuitry, should reduce tag costs to a few pence (or cents).

Auto-ID has a standardised 96 bit key or Electronic Product Code, that includes version number, manufacturer, serial number and/or product type. CDAC has set up a demonstration robotic warehouse cell with support from Gillette, Omron, Fanuc and Agent Oriented Software that uses Auto-ID and agent technology to select groups of products and load them into boxes for delivery. The agent technology allows the system to immediately adapt itself to jams and breakdowns, and also to prioritise deliveries so as to maximise profits.

Honeywell has been undertaking research into using agents for dynamic supply chain management. According to Technology Director Andrew Ogden-Swift, agents can be situated at all nodes in the chain. Those attached to suppliers are concerned with what suppliers can do, while those attached to customers relate to priorities - who will accept late deliveries with what penalties - and profits.

Ogden Smith also explained that his company, and others engaged in process control, while not using agents, are already moving towards something very similar, but from a different direction. Model based optimising controllers of one kind and another are widely used in the process industries, particularly in oil refineries. In boiler systems, multiple controllers on separate process units are co-ordinated by a multi-unit optimiser, which means that the individual units are to some extent autonomous, and interact with each other, even if they do not negotiate with each other directly.

Apart from the supply chain management project, most of Honeywell's agent development projects seem to be funded by DARPA, the US Defense Advanced Research Projects Agency. A number of these are battle management related. These have relevance to business because while oil companies may not physically destroy each other's installations, they do do their best to destroy competitor businesses by other means.

With UAVs, or as Dr Jeremy Baxter of Qinetiq explained, with tanks, the technique is to attach agents to each vehicle and group of vehicles. The vehicle agents are programmed so they achieve certain objectives, but in the event of some vehicles being unable to fulfil their initially assigned tasks, others may have their tasks assigned to them in order that overall goals may still be attained, if possible. In business management terms, this might mean workers and managers being automatically assigned new tasks and responsibilities in the event of others falling sick or leaving. Lyndon Lee, from BT described a proposal to assign agents to enquiries made to call centres, to negotiate with agents assigned to personnel and departments to direct enquiries to the personnel best able to deal with them. In the event of customers deciding to purchase services, agents would route the order by a dynamically optimised path through the whole process of checking credit, and finding resource to deliver.

Lee envisaged that if the enquiry or order is web based, the computer system graphically show the user exactly how it was being processed and where it had got to at any one time, in a manner not dissimilar to Professor Marik's simulation of a self optimising conveyor system. Other commercial companies known to be seriously investigating the implementation of agent technology include Daimler Chrysler and various Japanese car companies. McFarlane expects to see partial system usage in industry in, "The near future", with full agent usage in elements of the supply chain in the next three years.

Centre for Distributed Automation and Control
Professor Vladimir Marik at Rockwell
Andrew Ogden-Swift at Honeywell


* Software agents can be attached to orders, products and the resources used to manufacture them
* Agents act autonomously and negotiate with each other
* Agents can also advise other agents of events they should be made aware of on a need to know basis

For more technical developments see

Back to list of stories