Offline Neural Architecture Search for Time Series Forecasting

• NeuroEvolution Neural Architecture Search to Design Efficient RNNs for Time Series Forecasting (Implemented on EXAMM)
• Efficient time series forecasting can help make informed decisions in various industries, such as finance, energy, and manufacturing.
• Distributed population repopulation for efficient search
• Integrated into coal fired power plant management system and saved $7.3 million on revenue loss
• Evolved lightweight RNNs that can achieve state-of-the-art Transformer performance for stock forecasting
• Neural architecture search designed RNNs can reach samiliar performance as attention-based Transformer models
• An example of EXAMM generated RNN for stock forecasting
• Call for collaboration: time series analysis on edge devices and on other real-world applications

Online Neural Architecture Search for Time Series Forecasting

• The first online neural architecture search for real time RNN evolution and training.
• RNNs are dynamically generated and trained online without requiring pre-training.
• The models adapt to perform real-time predictions on datasets with data shift.
• Successfully deployed on raw wind turbine engine data streams.
• ONE-NAS Flowchart

Classification -- Determine Flight Phases of General Aviation Data

• A lot of real-world datsets are unlabeled, labeling them can be time and cost expensive.
• Real-world, large-scale, unlabeled data
• Class imbalance problem
• Use minimum number of labels for phase of flight classification
• Minimally Supervised Learning with Self Organizing Maps (MS-SOM) -- Classification

Regression -- Coal Property Estimation

• To require coal property values, coal samples need to be collected and tested in lab. Which is expensive and hard to get during COVID.
• This is novel approach as most of the minimally supervised learning problems are classification problems.
• Minimally Supervised Learning with Self Organizing Maps (MS-SOM) -- Regression