Distributed contract networks of sensor agents with adaptive reconfiguration: modelling, simulation, implementation and experiments

As both sensors and communication technologies are becoming more powerful at decreasing costs, sensor networks are constantly growing in complexity, i.e. in the information they generate and in the need for coordinating network resources. In this paper we comparatively explore the performance of multi-level hierarchical networks and flat networks, i.e. arrays of sensor agents with specific competences but without predefined communication or control structures. We develop a model of the performance of these types of networks for analysing the throughput of sensing tasks depending on a variety of network and sensor parameters. This analytical model serves as the basis for numerical simulation runs, which reveal the specific advantages of the different network structures. Motivated by these results, we develop an adaptive control scheme, which enables the network to organise itself at run-time so as to achieve the highest result precision in minimum time with minimum communication overhead, maximum parallelism and maximum fault tolerance. This control scheme is a computer-operational version of the contract network, originally proposed for problem solving in distributed artificial intelligence, with agents teaming up dynamically after a negotiation phase following a task specification. The metaphor of contracting also holds the potential to integrate agents that control actuators, e.g. cooperating robots, by assigning appropriate competences to them. Hence the design of sensorimotor systems, whose behaviour is adaptively controlled in a situation-dependent way by a team of different sensors, is a straightforward extension. Moreover, it is also easy to extend the network by adding agents that do not have a sensing component and can only perform sensor data processing tasks. We introduce a complete software framework for implementing these kinds of networks. It consists of a formal specification of the requirements that the sensory component of an agent has to meet and a comprehensive library of C++ objects encapsulating all of the negotiation protocol and communications routines. In the experimental section of the paper we show how several uncalibrated cameras are used to estimate the trajectory of a manipulator in order to guide the manipulator towards a target in real time. We also illustrate how agent-teams may easily reconfigure during task execution in the case of unexpected events. We then go on to show how complex vision tasks can be broken down into tasks that identical agents may work on in parallel. Finally we show how fusion for vision at a symbolic, i.e. object description level, may contribute to solve the hard problem of recognising occluded objects from different view points. We conclude the paper by summarising our results and sketching future research, in particular the automatic mutual reprogramming of agents using a specifically developed ontology.