梯度策略的多目标GANs帕累托最优解算法

Multi-objective GANs Pareto Optimality Algorithm Using Gradient Strategy

针对基于梯度策略的多目标优化算法无法适用于多目标、高维度的生成对抗网络(Generative Adversarial Nets, GANs)及多目标GANs中利用交叉验证产生次优解,极难求得最优解等问题,提出一种基于梯度策略的多目标GANs帕累托最优解算法。该算法采用硬参数共享方式,将多目标优化分解为多个两目标优化,确定多目标权重参数后,沿着梯度方向进行线性搜索,最终确定帕累托最优解。理论上,在弱条件约束下,证明了所提算法能够确切地产生帕累托最优解。实验上,将所提算法应用到图像处理的常见领域,对比所提算法与原算法的性能。结果表明,当目标数量大于2时,所提算法能够产生明显的性能优势。

To address the problems that the gradient-based multi-objective algorithm cannot be applicable for multi-objective Generative Adversarial Nets(GANs)with high dimension and multiple tasks. Then, considering a fact that solutions with cross-of-validation can be sub-optimal and it is hard to search global solutions in optimizing multi-objective GANs, this paper presents a multi-objective GANs Pareto optimality algorithm based on gradient strategy. The proposed algorithm uses hardware parameter sharing method and decomposes multi-objective optimization into multiple binary-objective optimizations. Then, it computes and determines all the weighted parameters and searches Pareto optimality along the gradient direction, which can yield exactly Pareto optimality. Theoretically, it has been proved that the proposed method can result in one Pareto optimality with detailed demonstration. Practically, the proposed algorithm has been applied into common sub-fields of image processing to compare the source and proposed algorithm in the same setting. The experimental results show that the proposed algorithm has outperformed than source algorithms when the number of tasks is over 2.