Gaussian Process for Prediction

We are developing computationally efficient online formulations of multiple output Gaussian process and applying them to sensor network data. The Gaussian process learns delays and correlations between sensors to constructs a probabilistic model that can be used to predict missing sensor values, make short term predictions into the future, or perform adaptive sampling (taking the minimum number to ensure that uncertainty is maintained below a threshold value). A demonstration of this algorithm applied to tide sensor data from a network of weather stations on the south coast of England is available here. A paper describing the algorithm in more detail can be found here.

Consistent Multiple Hypothesis Estimation with Faulty Sensors

We are developing robust fusion algorithms based on generalized covariance union (GCU) that allows consistent multiple hypothesis estimation in the face of faulty or untrustworthy sensors, or incomplete probabilistic models. A demonstration of this algorithm applied to the problem of inferring the temperature of surrounding buildings from a network of sensors (some of which are faulty) deployed within a city (based on the Robocup Rescue Simulation Environment) is available here.

Variable Selection in Streaming Data

We are developing efficient algorithms to perform online selection of variables from streaming data. Such algorithms have application in the problem of selecting readings from a subset of sensors to ensure the best possible prediction of future measurements, whilst minimizing bandwidth consumption. A demonstration of a novel algorithm (that extends the Lasso algorithm to incorporate reinforcement learning) is shown here, where it is used to select in real time, which subset of air temperature sensors to request a reading from, in order to achieve the best possible prediction.

Bidding Strategies for Efficient Resource Allocation

We are developing buyer and seller strategies for online auctions that permit the efficient allocation of goods and tasks in environments where agents are selfish and rational. These strategies allow agents to maximise their utility when bidding in multiple auctions at the same time. The project has generated a number of papers on the subject and these can be found on the Publications page.

The auction demos feature a number of ambulance centers that are responsible for rescueing trapped civilians in different parts of the map. We consider three different cases. Click on the link to see a video of each case:

• No Auctions: each center has a limited number of ambulances to save the civilians. The centers cannot share their resources since there exists no mechanism to do so dynamically in the case of major disasters..

• Dynamic Auctions: the ambulances are maintained by one central authority. Whenever centers feel the need for a number of ambulances, they bid for the required number of ambulances according to the criticality of the civilians' health. As soon as the proposed rescue task is completed by the hired ambulance, the latter is returned to the central authority which re-auctions it to prospective bidding centers.

• Distributed Auctions: each center has a number of ambulances to perform a number of rescue tasks. However, some of them may require more ambulances than they currently have and others may require less. Then, those excess ambulances are auctioned off and centers that need ambulances bid for these. The ambulances are allocated to the highest bidder (i.e. relative to the criticality of the civilian to be saved). When the ambulance has finished its task it is returned to its owner. Moreover, the demo shows that if a center fails, the ambulances can easily be re-allocated to the adjacent center and the dynamic allocation can be performed as before.

Anytime Coalition Formation

In order to coordinate a number of agents (e.g. ambulances or fire brigades) efficiently, it is important to distribute them in a number of efficient teams coalitions. Given $N$ agents, the number of coalitions that could be generated is 2^N. Moreover, each coalition may have a value which represents how efficiently it can perform tasks. This may be due to differences in capabilities or constraints. Now, it is important to select the best coalitions and this problem is termed the coalition structure generation problem. We have developed the fastest algorithm for this purpose and a paper describing our algorithm can be found here. To demonstrate the coalition structure generation process in our simulator, we show how a number of ambulances owned by a single center can split up into a number of efficient coalitions in order to rescue a number of civilians in the environment. Click here for a video.

Adaptive on-line decision algorithms for building evacuation

In this scenario we demonstrate the use of smart on-line algorithms that provide directions regarding the best available path towards an exit during a building evacuation in the presence of a hazard. The technical approach is based on a network of sensor and decision nodes positioned in specific locations. The paths that can be followed by an evacuee or by emergency personnel can be linked to progress along the decision and sensor nodes. Each sensor node is location aware and able to sense information regarding the environment, such as hazard intensity and location.Such algorithms can be used in practice with either “visible panel” indicators in the building which are controlled by the path finding algorithm, or via data exchanged between wireless nodes and portable communication devices (such as a PDAs) carried by the evacuees and the emergency personnel. Click here for a video.

Optimisation techniques for allocation of rescuers in disaster areas

Distributed decision making with limited communication implies that dinstinct rescuers must act without coordination during the decision process. We have developed a novel algorithmic approach based on designing an "oracle" that uses a neural network for mapping input information into allocation of destinations for the rescuers. The specific neural network model is the RNN with synchronised interactions which we train with situations that are similar to those encountered during an emergency. In the demonstrated scenario, the rescuers have an initial, incorrect estimation of the number of injured civilians. When a rescuer visits an injury location, he informs the others about the correct number of injured civilians and then each rescuer individually uses his "oracle" to decide whether to change or not destination. Click here for a video.