# Parameter Estimation in Abruptly Changing Dynamic Environments

## TL;DR

This paper introduces a novel online method for tracking statistical properties of data streams with abrupt changes, combining an estimator suitable for stationary data with an event detection based on the difference between estimators, outperforming previous approaches.

## Contribution

We develop a theoretically grounded detection procedure using the difference between a stationary estimator and a Stochastic Learning Weak Estimator, improving change detection in non-stationary environments.

## Key findings

- Effective in detecting abrupt changes in data streams
- Performs well on synthetic and real-world news classification data
- Easier to tune compared to existing methods

## Abstract

Many real-life dynamical systems change abruptly followed by almost stationary periods. In this paper, we consider streams of data with such abrupt behavior and investigate the problem of tracking their statistical properties in an online manner. We devise a tracking procedure where an estimator that is suitable for a stationary environment is combined together with an event detection method such that the estimator rapidly can jump to a more suitable value if an event is detected. Combining an estimation procedure with detection procedure is commonly known idea in the literature. However, our contribution lies in building the detection procedure based on the difference between the stationary estimator and a Stochastic Learning Weak Estimator (SLWE). The SLWE estimator is known to be the state-of-the art approach to tracking properties of non-stationary environments and thus should be a better choice to detect changes in abruptly changing environments than the far more common sliding window based approaches. To the best of our knowledge, the event detection procedure suggested by Ross et al. (2012) is the only procedure in the literature taking advantage of the powerful tracking properties of the SLWE estimator. The procedure in Ross et al. is however quite complex and not well founded theoretically compared to the procedures in this paper. In this paper, we focus on estimation procedure for the binomial and multinomial distributions, but our approach can be easily generalized to cover other distributions as well. Extensive simulation results based on both synthetic and real-life data related to news classification demonstrate that our estimation procedure is easy to tune and performs well.