Several model classes for system evaluation exist and are capable of representing the kind of complex behavior exhibited by contemporary distributed computer architectures and computer-communication networks. However, a number of problems associated with the evaluation of large-scale systems arise because of the size and complexity of the stochastic process derived from the underlying net model, which serves as a 'base model' for subsequent solution of the measures in question. If this base model is constructed by standard means, e.g., it is identified with the marking behavior of the net, traditional methods of solution quickly become intractable for large systems, limiting their application to systems of only moderate complexity. This problem is addressed in the stochastic activity networks (SANs) by developing base model construction methods that account for symmetries in SAN structure and are tailored to the variable in question (e.g., response time, time to failure, etc.). It is found that such a technique can yield dramatic reductions in state-space size while preserving stochastic properties required for practical means of solution. This technique permits direct construction of a reduced base model, thus avoiding size limitations associated with more traditional approaches to model amplification.