Stochastic token-generator models for simulating industrial message logs

Wilts, Stevan (2020) Stochastic token-generator models for simulating industrial message logs. Bachelor's Thesis, Artificial Intelligence.

Preview	Text Bachelors_project_Stevan_Wilts.pdf Download (3MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (97kB)

Abstract

Many machine learning implementations rely on synthetic data, but there are few, if any, general solutions for generating data. This study explores a new method for stochastic token-generation in the context of industrial message logs, using finite-state machines. For each of three different data distributions 50.000 finite-state machines are created. These machines can generate token sequences of arbitrary lengths and are randomly generated within a bounded parameter space. The distributions of the generated token sequences are compared to the target distributions and the machine that generates the sequence with the smallest mean squared error to the target distribution is selected. To test the predictability of the selected datasets, four different neural networks are trained on each set. Each of these models performed equally with 80 – 92% accuracy, depending on the dataset, and only three percent better than they would perform by only repeating the last message code. This shows that there is not enough internal dependency in the data to make reliable predictions about the future. This technique can however be used to generate sequences that can be used in the field of machine-learning to learn to solve these specific sequential problems.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Schomaker, L.R.B.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	03 Feb 2020 12:30
Last Modified:	03 Feb 2020 12:30
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/21493

Actions (login required)

View Item