Extended Trust Tech Methodology For Nonlinear Optimization

Many theoretical and practical problems can be formulated as a global optimization problem. Traditional local optimization methods can only attain a local optimal solution and be entrapped in the local optimal solution; while existing global optimization algorithms usually sparsely approximates the global optimal solution in a stochastic manner. In contrast, the transformation under stability-retaining equilibrium characterization (TRUST-TECH) methodology prevails over existing algorithms due to its capability of locating multiple, if not all, local optimal solutions to the optimization problem deterministically and systematically in a tier-by-tier manner. The TRUST-TECH methodology was developed to solve unconstrained and constrained nonlinear optimization problems. This work extends the TRUST-TECH methodology by incorporating new analytical results, developing new solution methods and solving new problems in practical applications. This work first provides analytical results regarding the invariance of partial stability region in quasi-gradient systems. Our motivation is to resolve numerical difficulties arising in implementations of trajectory based methods, including TRUST-TECH. Improved algorithms were developed to resolve these issues by altering the original problem to speed-up movement of the trajectory. However, such operations can lead the trajectory converge to a different solution, which could be undesired under specific situations. This work attempts to answer the question regarding invariant convergence for a special class of numerical operations whose dynamical behaviours can be characterized by a quasi-gradient dynamical system. To this end, we study relationship between a gradient dynamical system and its associated quasi-gradient system and reveal the invariance of partial stability region in the quasi-gradient system. These analytical results lead to methods for checking invariant convergence of the trajectory starting from a given point in the quasi-gradient system and the algorithm to maintain invariant convergence. This work also develops new solution methods to enhance TRUST-TECH's capability of solving constrained nonlinear optimization problems and applies them to solve practical problems arising in different applications. Specifically, TRUST-TECH based methods are first developed for feasibility computation and restoration and are applied to power system applications, including power flow computation and feasibility restoration for infeasible optimal power flow problems. Indeed, a unified framework based on TRUST-TECH is introduced for analysing feasibility and infeasibility for nonlinear problems. Secondly, the TRUST-TECH based interior point method (TT-IPM) and the reduced projected gradient method are developed to better tackle constrained nonlinear optimization problems. As application, the TT-IPM method is used to solve mixed-integer nonlinear programs (MINLPs). Finally, this work develops the ensemble of optimal, input-pruned neural networks using TRUST-TECH (ELITE) method for constructing high-quality neural network ensembles and applies ELITE to build a short-term load forecaster named ELITE-STLF with promising performance. Possible extensions of the TRUST-TECH methodology to a much broader range of optimization models, including multi-objective optimization and variational optimization, are suggested for future research efforts.