TY  - CONF
T1  - Reliable learning-based controllers and how structured simulation is a path towards them
A1  - Kušić, Krešimir
A1  - Schumann, René
A1  - Gregurić, Martin
A1  - Ivanjko, Edouard
A1  - Šoštarić, Marko
TI  - Proceedings of 5th International Conference on Advances in Signal Processing and Artificial Intelligence (ASPAI’ 2023)
Y1  - 2023
SP  - 268
EP  - 274
CY  - TENERIFE (CANARY ISLANDS), SPAIN
KW  - Controller reliability
KW  - Learning-based controller
KW  - structured simulation
KW  - Variable speed limit
N2  - New approaches to control stochastic non-linear time-variant processes include the application of machine learning techniques. One of the problems with learning- based controllers is their reliability in a wide area of process parameters as the controller is trained using a limited set of representative scenarios, either chosen by the designer or taken from historic records. Thus, reliable controller behavior can be guaranteed only in scenarios applied during controller training. Due to the very larger number of random variables and possible scenarios, not all variations can be applied in the controller training process using simulators to guarantee good controller behavior when applied in a real system. One case is traffic control (signal programs, variable speed limit, ramp metering) having large travel patterns variety. The concept of Structured Simulations Framework (SSF) can cover most probable learning scenarios. Thus, applying SSF enables a systematic controller training approach by complementing existing scenarios with synthesized ones that evoke or replicate substantial aspects of real traffic. Such training is necessary to ensure reliable learning-based controllers. This paper discusses the concept of applying SSF to ensure the reliability of learning-based controllers and proposes the application in traffic control for the case of variable speed limits on motorways.
ER  -