|
EPSRC Network on Evolvability in Biology & Software SystemsSoftware Evolution and Evolutionary Computation Symposium Abstracts
University of Hertfordshire, Hatfield, U.K.
|
IAN W. MARSHALL, CHRIS ROADKNIGHT AND ANTONIO GONZALEZ
BT Exact Technologies, Adastral Park, Ipswich, IP5 3RE, U.K.
ian.w.marshall@bt.com, christopher.roadknight@bt.com
Extrapolation of current trends for ownership of microprocessors [1] suggests that within 10 years it is possible that many individuals will own in excess of a thousand microcontrollers. If, as seems increasingly likely, pervasive computing on this scale is realized, users will be faced with a major investment of time and money in configuration and maintenance activities. To minimize this impact it is important to investigate ways of automating low level management processes and enabling many pervasive computing devices to operate almost autonomously [2]. At the same time it is vital to ensure that the management processes are able to adapt to new requirements and applications, since it is likely that developments will be extremely rapid and unpredictable [3]. These requirements suggest an evolutionary solution would be ideal.
In our effort to address the problems we have been investigating the possibility of linking an accretionary management solution (policy based management [4]) with an evolutionary algorithm derived from prokaryote biology [5]. The algorithm combines rapid learning (plasmid exchange), automated configuration (limited motility of individuals) and evolutionary learning via a conventional mutation based GA. Simulations of up to 5000 nodes have shown the algorithm is able to create and maintain a system that performs at least as well as conventional designs, whilst requiring almost no human input or intervention. To provide a practical verification of the simulation results we are currently embodying the algorithm in an ad-hoc network consisting of 20 nodes, where the nodes are intended to emulate distributed sensor controllers. Each node consists of 3 sensors, 3 actuators, a 16 bit microprocessor and an infra-red transceiver.
A key aspect of this implementation is a novel operating system that supports dynamic extension of basic functionality through addition of policies (plasmids). The system can autonomously flush accreted functions that prove to be detrimental, and attempt to acquire a different version of the desired function that may perform better. This will enable adaptation to occur in the kernel and in the application space simultaneously, and should be a closer analogue of the multilayer adaptation observed in real biology than many previous experiments. We will present some preliminary results that demonstrate the benefits of this approach, and show how it can be used to deliver customized secure communication in addition to the application management we have demonstrated previously.