Simulation of human behavior in a virtual living environment : for the generation of statistical validation data
| Autor: | Arbeiter, Maximilian |
|---|---|
| Jazyk: | angličtina |
| Předmět: | |
| Popis: | Today we live in an age of big data. Yet for many applications, it is still not always trivial or too costly to collect a significant amount of statistical validation data. One such situation is the verification of an anomaly detection system. This system should be installed into the living environment of lonely living elderly persons. Based on several sensor measurements, it should be able to learn the persons behavior and to detect exceptions. Collecting such validation data would require a huge amount of test persons. Fortunately, there is an alternative: Generating test data by simulation. It is the goal of this dissertation to describe and create a simulation environment for the generation of the required sensor data.The core of this simulator is the simulation of human behavior consisting of several parts which are discussed in this dissertation. Human behavior can be described as a sequence of activities. Each activity consists of several actions. To get all required actions to perform a certain activity, an automated planning algorithm is needed. Here, so-called stochastic behavior trees are used to find such a plan. Additionally, the problem to choose between various different plans leading to the same goal is also considered. Using some arguments from classical decision theory, it is shown how the agent has to select such an optimal plan.For the simulation of human behavior, it is mainly necessary to select reasonable activities. These selection problems are considered using decision theory. Besides the classical maximum utility principle, there are several alternatives. In this dissertation, some of them are combined to get a very general and flexible decision rule which takes past cases, expected outcomes and observed data into account.Further, multi-stage decisions are considered, i.e., the selection of a sequence of activities. This leads to Markov decision processes. Here, processes with a high dimensional state space are investigated. Such decision processes need special simulation algorithms to be solved efficiently. Two such algorithms are presented and some of their mathematical-statistical properties are analyzed.Finally, the simulated data are used to perform activity recognition. This is done to check if the simulators virtual sensors provide enough information to detect exceptions.Here, Bayesian multinomial regression is used for this task. This analysis shows that the chosen sensors are sufficient for a relatively precise classification of activities. Maximilian Arbeiter Dissertation Alpen-Adria-Universität Klagenfurt 2021 |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|