A STUDY ON POTENTIAL OF GENETIC ALGORITHMS IN CIVILENGINEERING

Abstract

Genetic algorithms (GAs), which are search algorithms based on the mechanics of natural selection, have received a lot of attention in various disciplines of engineering and sciences, With this in view, GA application to certain problems from Civil Engineering has been addressed in the present study. The literature regarding different aspects ofGAs and their application is surveyed and critically reviewed. The difficulty in their implementation and other gaps are identified. In reality, problems may vary in complexity a lot. As it is difficult to have consensus on the degree of complexity of a given problem, investigation is started from a simple possible case to evaluate the performance of GAs. Towards this, the problem of simple unconstrained linear function is investigated using exhaustive range of GA parameters and with three selection strategies for tournament selection scheme. The GA's performance is judged from the numbers of function evaluation and the number of generations yielding the optimum solution. The efficiency of GA is observed to be improving by relaxing function evaluations in case individuals (solution to optimisation problem) escape from variational operators of GA. The performance of parameter-less GA is also evaluated by testing it again on a same simple function, and observations are reported. In parameter-less GA, the probability of cross-over and setting selection rate is taken outside from user's purview. A number of GAs run are used simultaneously with different population sizes. It has been observed that as parameter-less GA progresses, smaller population are removed and larger populations are inducted. It is found that parameter-less GA is not very successful in term of computation. The GA parameters, which are found from simple linear problem, are tried on problem of trusses (planer and space), which represent a class of problems with linear objective function having linear as well as non-linear constraints. The constraint with regards to axial stress and displacement are considered. The performance of GA parameters is evaluated by increasing the complexity of problem in terms of m number of design variables. Indeterminate truss and space frame are also used while judging performance of GAs. The cost is taken as objective function in case of problem of non-linear objec tive function with constraints. This class of problem is represented by design of reinforced concrete beam section. The problem is almost impossible to be solved by conventional derivative based methods, as design variables which govern the de tailing of reinforcement, are considered. The sensitivity results obtained by GA is studied with respect to probabilities of cross-over and mutation. The effect of selection pressure is also studied. The effect of penalty parameter on the quality of solution is examined by taking different values of penalty parameter are used to transform constrained optimisation problem to unconstrained optimisation problem. The comparative study is also performed between cost and weight optimisation. The network analysis problem is used to test suitability of GAs for analysis problem. Here, the problem is also having continuous design variables. So, the working of GA is also examined for optimisation problem with continuous decision variables. The objective function is taken as error function, the value of which is minimised. The problem is solved using Hazen-William expression and Darcy- Weisbach's formula. The comparative study is performed with commonly used other traditional methods. To evaluate GA's behaviour on various types of problems, it is also used in prob lems with noisy information (data) at different locations. The problem of estimation of characteristic values of parameters of an aquifer from discharge of artesian well is dealt with. Different level of noise is used to study its effect. The multi-location noise problem which is taken in this study is parameter estimation of Muskingum model for flood routing problem. The noise may be either in inflow or in outflow. Various combinations of noiseless and noise at different locations are considered while studying GA's working on noisy systems. GA parameter's sensitivity is ob served with different noise conditions. Results are compared with results obtained from conventional method (Newton Raphson). Further, GA is applied to multiobjective optimisation, where again space frame is optimised. The importance of selection of proper objectives and choice of proper range of design variables is emphasised. It is found that use of partial information from in-feasible solution may make GA based on binary representation to surpass GA with floating point representation in term of solution quality, which is reported earlier to perform well. To summarise, based on the application of GAs to certain problems of Civil Engineering, it can be inferred that GA has potential for application to a variety of IV CIVILENGINEERING problems. The use of parameter-less GA is, however, not observed to be very successful and attractive. For simple problems, it is possible to achieve the optimum solution at value of mutation probability lying within a range, which may be totally different for achieving the solution of a complex problem. The cross over probability is not very critical for good results, while mutation rate has been observed to be critical for success of GAs and appears to be problem specific. For parameter estimation problems, GAs are found to be more useful, when the system input and output contain noise level of varying extent. The elitism is found to be playing positive role consistently in all problems studied herein