Increased complexity evolution applied to evolvable hardware

Evolvable Hardware (EHW) has been proposed as a new method for designing systems for complex real world applications. One of the problems has been that only small and simple systems have been evolvable. This paper highlights some of the reasons that can explain why EHW has not yet been widely applied. Further, to make EHW more applicable, an increased complexity scheme is proposed, where a system is evolved by evolving smaller sub-systems. Experiments show that the number of generations required for evolution by the new method can be substantially reduced compared to evolving a system directly. This is with no lack of performance in the final system. Another experiment shows that gate level EHW lacks noise robustness. INTRODUCTION Evolvable hardware (EHW) has recently been introduced as a new scheme for designing systems for real world applications. So far the number of applications is highly limited. One of the main problems in evolving hardware systems seems to be the limitation in the chromosome string length [7]. A long string is required for representing a complex system. However, a larger number of generations are required by genetic algorithms (GA) as the string increases. Thus, work has been undertaken to try to diminish this limitation. Various experiments on speeding up the GA computation have been undertaken [2]. The schemes involve fitness computation in parallel or a partitioned population evolved in parallel. Experiments are based on speeding up the GA computation, rather than dividing the application into subtasks. This approach requires that GA finds a solution if it is allowed to compute enough generations. When very small applications require weeks of evolution time [12], there would probably be strict limitations on the systems evolvable even by parallel GA. Other approaches to the problem have been by using variable length chromosome [4]. Another option, called function level evolution, is to evolve at a higher level than gate level [9]. Most work is based on fixed functions. However, there has been work in Genetic Programming for evolving the functions [6]. The method is called Automatically Defined Functions (ADF) and is used in software evolution. Both gate and function level of EHW evolution have been applied to real applications. To control an artificial hand, a gate-level EHW chip has been designed [5]. The GA operations are undertaken in a separate unit within the chip. Simulations of data compression using function level evolution indicate performance comparable to other compression methods like JPEG compression [11]. The scheme is designed for implementation in a custom ASIC device. A function based FPGA has been proposed for applications like ATM cell scheduling [8] and adaptive equalizer in digital mobile communication [10]. Except for the limited number of real problems studied, there is a larger range of small and artificial problems, see [13, 15] The idea of gradual increase in complexity has been introduced for AI systems using software models [1]. In the work it is stated that humans undergo a process of development where we are able to perform more difficult tasks in more complex environments en route to the adult state. It is foreseen that the construction of such an AI system could scale autonomously. Which learning structures and organizational principles to apply is still an open question.