Learning Neural Networks That Can Sort
MetadataShow full item record
This thesis analyzes how neural networks can learn parallel sorting algorithms such as bitonic sorting networks. We discussed how neural networks perform at sorting when given no information or constraints about the allowable operations. We focused on analyzing how the architecture, training data, and length of the array impacted the neural network’s performance at sorting. After encountering challenges with using neural networks to sort, we analyzed how neural networks learn the building blocks for sorting (comparator and swap- ping operators). Once we saw that these basic operations cannot be learned, we framed parallel sorting as a Reinforcement Learning problem. Using Reinforcement Learning, we were able to learn parallel sorting algorithms for sequences of lengths 4 and 8 under certain conditions, specifically limiting the allowable actions. We concluded that using Deep Reinforcement Learning there is potential to learn parallel sorting algorithms without any constraints.