Table 1 Comparison between our proposed methods and some general optimization algorithms
From: Gumbel-softmax-based optimization: a simple general framework for optimization problems on graphs
Optimization algorithms | Pros | Cons |
|---|---|---|
Genetic algorithm | Population-based, easily implemented, commonly used on various problems | Do not use gradient information, do not scale well with complexity, unable to deal with constraints effectively |
Simulated annealing | Suitable for optimization problems where search space is discrete, unlikely stuck with local optimum | Do not use gradient information, slow convergence, limited to system size up to thousand |
Greedy | Fast and easily implemented, results are usually satisfying | It may fail to find global optimal on certain problems |
GSO | Taking advantage of gradient information using relaxation techniques, with automatic differentiation technique, it is much more faster than classic algorithms | There is still room for improvement of the solution, because it may stuck with local optimum sometimes |
EvoGSO | Both gradient and population-based algorithm, with evolution strategy, it is able to overcome local optimum to some extent and obtain better solutions | Usually more time consumption than original GSO in exchange for better results |