Gaussian Processes for ordinal regression

点击翻译按钮获取中文摘要

Problem or ChallengeHighly complex infrastructure projects rarely run to schedule as a result of human optimisation bias during planning, complex interactions between project activities, and failure to learn systematically from past projects. nPlan has amassed the largest proprietary collection of schedules data in the World and developed machine learning modelling technologies that train on historical data to predict future risk. In the real world however, the utility of trained models depends on evaluating well-calibrated probability distributions for predicting future outcomes. The aim of this project is to develop novel probabilistic models that optimise computation of these probability distributions by delineating aleatoric and epistemic contributions to uncertainty using Bayesian inference.MRes/PhD Project Objectives1. Develop graph and/or language models for efficient numerical representation of schedule features.2. Identify representative data sources with similar graph and/or language features to evaluate probabilistic models.3. Develop probabilistic models that infer relationships between schedule features and outcomes.4. Optimise modelling algorithms to be scalable for commercial implementation.PhD Project DescriptionThe development of probabilistic models build upon existing state-of-the-art deep learning technologies already developed at nPlan Since the nPlan project depends on simulating outcomes of schedules based on trained models, it is important that predicted probability distributions provide realistic representations of uncertainty. Since conventional deep learning models do not explicitly characterise probabilities, the aim of the project is to develop novel models that do. Access will be given to the nPlan proprietary data set and the nPlan Core code-base for efficient data reading. The scope of the PhD project is purposefully broad to encourage exploration of different project data types and models, while providing guidance from the expertise within nPlan and the Prof. Girolami group.MRes ComponentThe MRes component includes objectives 1&2 listed above. Since the nPlan dataset is very large, preliminary prototype models trained on suitable subsets will be developed within the first year. This work includes identification of data sources with similar structures (Gantt charts represented as directed acyclic graphs) for initial development and evaluation of probabilistic models.