Network-augmented Evolvable Reconfigurable Architectures: A Novel Platform for Evolvable Hardware

摘要

The research presented in this thesis focuses on Evolvable Hardware, a circuit design technique that relies on biologically-inspired algorithms and reconfigurable hardware to realize innovative designs. Circuits evolved with these methodologies can adapt to highly dynamic environments and are able to reconfigure themselves around faults which would permanently disable canonically-designed hardware. Furthermore, the evolutionary algorithms employed often produce hardware which a conventional, human-driven, design flow would have excluded a priori. These unconventional circuits have, in some cases, resulted in lower power consumption and higher performance with respect to their conventional counterparts.;Recently introduced platforms, directly coupling high-performance processing systems with reconfigurable logic, have considerably expanded the range of possibilities that can be attained by evolvable hardware. Exploiting one family of such platforms, a network-enabled evolvable hardware platform was developed.;Creating such a system invariably requires the development of several interconnected components, starting from the low-level architecture up to the network interconnections. The result is a novel and intuitive network-enhanced evolvable reconfigurable architecture, developed on an open-source Linux-based operating system. Aside from the network-related improvements, the platform also offers a radically different perspective into evolvable hardware, which enables effortless integration of new features and considerably reduces the effort required to create new evolvable hardware designs.;The evolutionary performance of the developed platform was evaluated on a widespread family of circuits, image filters. The development of this benchmark application enabled verification of the capabilities of the architecture as well as validation of the proposed network-based enhancements for evolvable hardware. In-depth testing, using a wide variety of different configurations, demonstrated the superior performance of the network-based evolvable hardware compared to that of conventional evolvable hardware architectures.