Exploring Massively Parallel Models and Architectures for Efficient Computation of Evolutionary Algorithms
'Evolutionary algorithms' is the collective name for a group of relatively new stochastic search algorithms that have several unique and interesting properties.
However, as with all search algorithms, evolutionary algorithms also have disadvantages. One of the most troublesome is that they require very time-consuming computations in order to reach a solution. Even though the calculations of the algorithms are simple, they have to be repeated so many times that the algorithms often require several days or even months of calculations to reach a solution when run in software on a modern computer.
In this thesis we present a way to speed up these calculations by combining different improvements. First, we choose a distributed computation model, called the Diffusion Model, that makes it possible to utilize massively parallel computation resources and thereby drastically reduce the computation time.
Second, we choose linear machine code as a representation for the evolutionary algorithm. These segments of code are executed in parallel by specially designed processing elements, embedded as independent, parallel resources in the system.
Third, on the basis of extensive simulations using a custom made simulator that allowed us to investigate the design space of the Diffusion Model in great detail, we determine a recommended set-up for the distributed Diffusion Model. This set-up includes the topology, the size and shape of the neighbourhoods and the local selection algorithm, as well as the parameters for the distributed evolutionary algorithm.
As a final factor for speeding up the calculations, we demonstrate how a system based on the massively parallel Diffusion Model can be implemented in hardware using VLSI technology.
Besides being efficient, the approach is also flexible enough to be used in different application areas and scalable enough to be able to address larger problems by using more computational resources. By implementing the system using VLSI technology, it becomes both compact and power efficient, making it possible to integrate evolutionary algorithms in many applications.
distributed population models
symbolic function regression
linear machine code
time series forecasting