GACEM: Generalized Autoregressive Cross Entropy Method for Multi-Modal Black Box Constraint Satisfaction

In this work we present a new method of black-box optimization and constraint\nsatisfaction. Existing algorithms that have attempted to solve this problem are\nunable to consider multiple modes, and are not able to adapt to changes in\nenvironment dynamics. To address these issues, we developed a modified\nCross-Entropy Method (CEM) that uses a masked auto-regressive neural network\nfor modeling uniform distributions over the solution space. We train the model\nusing maximum entropy policy gradient methods from Reinforcement Learning. Our\nalgorithm is able to express complicated solution spaces, thus allowing it to\ntrack a variety of different solution regions. We empirically compare our\nalgorithm with variations of CEM, including one with a Gaussian prior with\nfixed variance, and demonstrate better performance in terms of: number of\ndiverse solutions, better mode discovery in multi-modal problems, and better\nsample efficiency in certain cases.\n