Aladdin

Dealing with an environmental disaster or a terrorist incident involves rapidly shifting scenarios, where information changes constantly, and is often conflicting, making it hard to get an effective response and ensure the safety of emergency services personnel. The University of Southampton’s electronics and computer science department is applying its expertise on autonomous agents to develop software programs that in such situations can interact robustly to maintain data and information systems.

ALADDIN (Autonomous Learning Agents for Decentralised Data and Information Networks) is a multi-million pound, five-year project funded by the aerospace and defence group BAE Systems and EPSRC, with Southampton as the lead partner, and involving Imperial, Oxford and Bristol Universities. “We’ve looked at two main scenarios: a city-wide urban response, where you might want to coordinate police, fire, and ambulance to respond to an incident or an evolving series of incidents; and sensors in the Solent, where we’re looking at making predictions and extrapolations using data from sensors, which are sensing tide height, wind speed, and weather-related things,” says Nick Jennings, professor of computer science at Southampton. “Weather follows a pattern, so what’s further down the coast is – by and large – going to move up the coast in a few hours time.”

By bringing together many different sources of data, some of which might be better than others, and making informed estimates about missing data, ALADDIN can develop predictive algorithms that tell how, for example, a fire might spread, based on patterns that have been seen before.
Jennings’s team is applying a variety of techniques from machine learning, artificial intelligence, and Bayesian reasoning in this project. It is also using ‘game theory’, a theory of competition beloved of economists that is stated in terms of gains and losses among opposing players. “Game theory doesn’t work very well on people, because they don’t tend to behave predictably or rationally – but it does work well on software agents,” Jennings explains. “We use incentives so that an agent gets rewarded for particular actions, which encourages them to behave in a particular way.” For instance, an incident commander may be tempted to overplay the severity of an incident in order to make sure that he or she receives an adequate number of ambulances or fire engines; but if everyone does this, allocation is not very efficient. The idea, then, is to put mechanisms in place that reward asking for the right amount of resource, and that punish requests for too much. This is tricky – not least because the systems involved are typically open, and anyone can join or add their piece of software or sensor into the network. Just like humans, software agents have to learn who is reliable and trustworthy – and who isn’t.

The ALADDIN programme is over half-way through, scheduled to end in October 2010. The core intellectual property developed so far is a series of coordination algorithms, which pull different inputs together to coordinate a response. Already these coordination algorithms, and others for reasoning about uncertainty, are being tried out within BAE business units involved with logistics and coordinating supply chains. In the long term, the aim is that the technologies will benefit all the UK’s emergency services.