Predictive simulation based on dynamic optimization using musculoskeletal models is a powerful approach for studying biomechanics of human gait. Predictive simulation can be used for a variety of applications from designing assistive devices to testing theories of motor controls. However, one of the challenges in formulating the predictive dynamic optimization problem is that the cost function, which represents the underlying goal of the walking task (e.g., minimal energy consumption) is generally unknown and is assumed a priori. While different studies used different cost functions, the qualities of the gaits with those cost functions were often not provided. Therefore, this dissertation evaluates and examines different cost function forms for dynamic simulation of human walking. The problem of the walking cost function determination was cast as a bilevel optimization, which was solved using a nested evolutionary approach. The results showed cost functions based on a weighted combination of muscle-based performance criteria (e.g., metabolic cost, muscle fatigue), gait smoothness, and gait stability led to better walking solutions compared to any cost functions only based on muscle performance criteria. Further evaluations of the walking cost functions were done in the simulation cases of human walking augmented with assistive devices such as prosthesis and exoskeleton. The simulations of augmented walking were comparable to the experimental results, which suggests the potential of using the simulation approach to address problems of finding assistive device design and control.