%0 Thesis %1 aydin2022optimierte %A Aydin, Suat %D 2022 %E Kässinger, Johannes %K persival pn7-1 simtech %T Optimierte dynamische Updatestrategien für verteilte Simulationen unter der Verwendung von Vorhersagemodellen %X The usage of large NN allow the calculation of very accurate results for simulations. However, their usage suit only limited for mobile devices, due to possibly insufficient computation power of the mobile devices and consequently higher electricity demand, which results in less battery life. In this thesis a distributed system was developed for the simulation of an arm movement, consisting of a client and a server. A high quality NN is used on the server, which gets the input values of an arm movement and calculates the corresponding points of the upper arm‘s muscle surface in high quality and outputs them as result. The forecasting methods are operated mirrored on the client as well as the server to forecast on both the future input values of the arm movement and the future upper arm points as output values. The extended damped Holt‘s method is used as forecasting method, which is a further development of the damped Holt‘s method, developed by BM22. This is a time-series forecasting method which uses the past history values to forecast the future values. A low quality NN is available on the client as backup-solution, which produces less accurate output values, if a solution computed by the high quality NN of the server is not present. Several update strategies were realised for using the forecasting methods, when and how the client and the server update each other. As part of the updates, the model data of the forecasting methods are sent, which allow the exact reconstruction of the forecasting methods, instead of the history values. Furthermore, a mechanism was developed, which is able to detect if the arm stands still and as a result lowers the communication. The distributed system‘s multi-objective consists of avoiding the local execution on the client as far as possible and reaching the highest possible quality of output values by using as less as possible communication processes and bandwidth. The distributed system is able to reduce the error by approx. 90% compared to the only execution of client‘s backup solution, depending on the chosen update strategy. The local execution is avoided there largely. The communication processes, average forecasting durations of the forecasting methods and required bandwidths of the distributed system which have taken place vary partially great between the different used update strategies and will be looked closely and compared with each other in the evaluation.

@mastersthesis{aydin2022optimierte, abstract = {The usage of large NN allow the calculation of very accurate results for simulations. However, their usage suit only limited for mobile devices, due to possibly insufficient computation power of the mobile devices and consequently higher electricity demand, which results in less battery life. In this thesis a distributed system was developed for the simulation of an arm movement, consisting of a client and a server. A high quality NN is used on the server, which gets the input values of an arm movement and calculates the corresponding points of the upper arm‘s muscle surface in high quality and outputs them as result. The forecasting methods are operated mirrored on the client as well as the server to forecast on both the future input values of the arm movement and the future upper arm points as output values. The extended damped Holt‘s method is used as forecasting method, which is a further development of the damped Holt‘s method, developed by [BM22]. This is a time-series forecasting method which uses the past history values to forecast the future values. A low quality NN is available on the client as backup-solution, which produces less accurate output values, if a solution computed by the high quality NN of the server is not present. Several update strategies were realised for using the forecasting methods, when and how the client and the server update each other. As part of the updates, the model data of the forecasting methods are sent, which allow the exact reconstruction of the forecasting methods, instead of the history values. Furthermore, a mechanism was developed, which is able to detect if the arm stands still and as a result lowers the communication. The distributed system‘s multi-objective consists of avoiding the local execution on the client as far as possible and reaching the highest possible quality of output values by using as less as possible communication processes and bandwidth. The distributed system is able to reduce the error by approx. 90% compared to the only execution of client‘s backup solution, depending on the chosen update strategy. The local execution is avoided there largely. The communication processes, average forecasting durations of the forecasting methods and required bandwidths of the distributed system which have taken place vary partially great between the different used update strategies and will be looked closely and compared with each other in the evaluation.}, added-at = {2023-02-14T17:17:03.000+0100}, author = {Aydin, Suat}, biburl = {https://puma.ub.uni-stuttgart.de/bibtex/24658d8236593a2f2980a8a1cd1319154/kaessijs}, editor = {Kässinger, Johannes}, interhash = {579aa50d78dafcebb68cb9d35aadad00}, intrahash = {4658d8236593a2f2980a8a1cd1319154}, keywords = {persival pn7-1 simtech}, timestamp = {2023-02-15T16:07:50.000+0100}, title = {Optimierte dynamische Updatestrategien für verteilte Simulationen unter der Verwendung von Vorhersagemodellen}, type = {{B.S.} thesis}, year = 2022 }