: Since recognition of parameters and constitutive models is often highly nonlinear and multimodal in the large parameter space, it is necessary to look for the global optimum method and its parallel realization. This paper reports genetic algorithms for recognizing unknown coefficients in mechanical model and connection weights of neural network model for rock mechanical problems, and genetic programming algorithms to recognize the structure of mathematical/mechanical models and topology of rock mechanical neural network models. Finally, as examples, recognition of mechanical parameters of rockmass related to permanent shiplock in Three Gorges Project using an integration of genetic algorithm, neural network, and finite element method is given. Recognition of constitutive model for a soft rock using an integration of genetic algorithm and genetic programming is also discussed. RsmVPC, a new visual parallel computing environment, is developed on Windows platform for parallel realization of the algorithm mentioned above. A complex rock mechanical task can thus be divided and solved by many personal computers at the same time, not only a personal computer. Therefore, the scale of computing can be large enough and the computing velocity can be fast enough to satisfy needs of a true problem.
Mechanical behaviors of rock masses are very complex, due to be affected by many factors such as geological structures, water, temperature, stress, chemical erosion, and excavation, etc.. In many cases, these factors interact each other. Recognition of parameters is an optimum problem of multi-parameter combination in large space. Contrastively, recognition of analysis or constitutive models is a more complicated optimum problem of highly multimodal and multi-parameter combination in large space (Sakurai, 1986). Now a question is arisen. How to obtain the optimum solutions, not local optimum solution, in global space for these two kinds of search problem? There are two points should be addressed. One is search method. We concern that it shall be a global optimum method. Another is search efficiency. We hope it can obtain this optimum solution as fast as possible. For the former, generally, there are three main types of search methods: calculus-based, enumerative, and random. Since material constitutive model is highly nonlinear and of the large parameter space and highly multimodal, calculus-based and enumerative techniques were discounted as being either not robust enough or not efficient enough to deal with this problem. Thus, the two most efficient algorithms are the genetic algorithm and the genetic programming technique (Goldberg, 1989). For the later, even if genetic algorithm and genetic programming are two kinds of non-numerical paralle computing methods, parallel realization of these algorithms is also needed to obtain higher efficiency.