Final Activity Report Summary - EVOLCIRCUIT (Evolutionary Probabilistic Circuit Design)
High level of modern electronic systems integration and rapid evolution of technological processes result in significant complication of design process of electronic circuits. Smart design and optimisation tools are required to manage the complexity of the problem. The project focussed on the application of the evolutionary probabilistic circuit design methodology proposed by H. Mühlenbein et al. in order to design and optimise mixed analog and digital signal circuits.
Deterministic approaches are usually used for circuit design and optimisation, requiring computationally expensive derivatives calculations. In addition, for some important practical design tasks derivatives calculation is difficult. It results in hard constraints for deterministic approaches' applications for circuit design and optimisation. On the contrary, evolutionary probabilistic models could be applied on design and optimisation problems without gradients and second-order derivatives' calculation. The evolutionary probabilistic circuit design and optimisation starts with a population of random design solutions. During evolutionary trials possible design solutions evolve based on the probability distribution. At the end of the evolving process the best design solutions are available.
The project was concentrated on the enhancement of the evolutionary probabilistic circuit design methodology for mixed analog-digital signal circuits. Mixed analog-digital circuits are widely used in different applications, e.g. automotive electronics, industrial electronics, etc. It should be noted that the circuit behaviour is described by nonlinear differential algebraic equations in a general case. Unfortunately, only numerical simulation is possible for mixed analog-digital signal circuits. This simulation is computationally extremely expensive. Therefore, a design and optimisation algorithm should be able to find good solutions very fast.
However, recommendations for the effective evolutionary probabilistic circuit design and optimisation of mixed analog-digital signal circuits were unclear. A choice of a fitness function evaluation schedule and a circuit representation for mixed analog-digital signal circuits was inexplicit as well. During the project the evolutionary probabilistic design and optimisation performance for mixed analog-digital signal circuits with different fitness functions' evaluation techniques and different circuit representations were investigated. It was shown that the evolutionary probabilistic circuit design and optimisation with higher population sizes seemed to be more suitable for multi-objective optimisation of mixed analog-digital signal circuits, while elitism could enhance the single-objective optimisation of mixed analog-digital signal circuits. It was also found that the behaviour of the multi-objective evolutionary probabilistic design and optimisation could change dramatically because of conflicts in the objectives to be optimised.
It was furthermore shown that the dynamic fitness function evaluation schedule seemed to be a good compromise between computational costs and optimisation efficiency, while the co-evolution fitness function evaluation schedule could require the smallest computational costs. It was also found that the binary branch-list circuit representation seemed to be the most effective to enhance optimisation efficiency. The methodology was validated by optimising the symmetry recognition circuit containing analogue-digital converter, flip-flops etc. and the custom cell.
According to our real world applications, the benchmark of our technology was the integrated 0.5 micron BiCMOS B6CA technology, developed by Infineon Technologies AG. Real world compact models of devices corresponding to the technology were used during evolutionary trials. Experimental results validated the methodology by comparing the performance of the optimised circuits with the initial circuits. This project generated innovative contribution to computer aided circuit design and evolutionary computation.
Deterministic approaches are usually used for circuit design and optimisation, requiring computationally expensive derivatives calculations. In addition, for some important practical design tasks derivatives calculation is difficult. It results in hard constraints for deterministic approaches' applications for circuit design and optimisation. On the contrary, evolutionary probabilistic models could be applied on design and optimisation problems without gradients and second-order derivatives' calculation. The evolutionary probabilistic circuit design and optimisation starts with a population of random design solutions. During evolutionary trials possible design solutions evolve based on the probability distribution. At the end of the evolving process the best design solutions are available.
The project was concentrated on the enhancement of the evolutionary probabilistic circuit design methodology for mixed analog-digital signal circuits. Mixed analog-digital circuits are widely used in different applications, e.g. automotive electronics, industrial electronics, etc. It should be noted that the circuit behaviour is described by nonlinear differential algebraic equations in a general case. Unfortunately, only numerical simulation is possible for mixed analog-digital signal circuits. This simulation is computationally extremely expensive. Therefore, a design and optimisation algorithm should be able to find good solutions very fast.
However, recommendations for the effective evolutionary probabilistic circuit design and optimisation of mixed analog-digital signal circuits were unclear. A choice of a fitness function evaluation schedule and a circuit representation for mixed analog-digital signal circuits was inexplicit as well. During the project the evolutionary probabilistic design and optimisation performance for mixed analog-digital signal circuits with different fitness functions' evaluation techniques and different circuit representations were investigated. It was shown that the evolutionary probabilistic circuit design and optimisation with higher population sizes seemed to be more suitable for multi-objective optimisation of mixed analog-digital signal circuits, while elitism could enhance the single-objective optimisation of mixed analog-digital signal circuits. It was also found that the behaviour of the multi-objective evolutionary probabilistic design and optimisation could change dramatically because of conflicts in the objectives to be optimised.
It was furthermore shown that the dynamic fitness function evaluation schedule seemed to be a good compromise between computational costs and optimisation efficiency, while the co-evolution fitness function evaluation schedule could require the smallest computational costs. It was also found that the binary branch-list circuit representation seemed to be the most effective to enhance optimisation efficiency. The methodology was validated by optimising the symmetry recognition circuit containing analogue-digital converter, flip-flops etc. and the custom cell.
According to our real world applications, the benchmark of our technology was the integrated 0.5 micron BiCMOS B6CA technology, developed by Infineon Technologies AG. Real world compact models of devices corresponding to the technology were used during evolutionary trials. Experimental results validated the methodology by comparing the performance of the optimised circuits with the initial circuits. This project generated innovative contribution to computer aided circuit design and evolutionary computation.