In this project, we aim to develop a time series analysis framework using modern machine learning techniques. The project focuses on solving the two fundamental tasks in time series analysis.
1. Multivariate Demand Forecasting With Uncertainty Estimate
Our approach is based on Bayesian LSTM. Besides direct forecasting applications, our research also studies time series analysis in the context of non-markovian reinforcement learning.
The most common framework for RL relies heavily on the markovian-observation assumption, that an agent’s next observation of the environment depends only on its current observation and next action. But most real-world environments violate this assumption, as when the agent’s observation is a lossy function of the true environment state. We call such environments partially observable. In farming for example, we can see the plant, but we do not have access to the full state of the plant or the surrounding environment. We are exploring using RNNs for multivariate time series to model the environment and the hidden state.
2. Anomaly Detection
2.1. Streaming Least Squares Algorithm for Univariate Time Series Anomaly Detection
Open source, SLS explained
2.2. Multivariate time series analysis for anomaly prediction