Stochastic Safety and Safe Decisions
Event details
| Date | 04.04.2025 |
| Hour | 11:00 › 12:00 |
| Speaker | Rafal Wisniewski, Professor and Deputy Head for research in the Department of Electronic Systems, Aalborg University, Denmark |
| Location | |
| Category | Conferences - Seminars |
| Event Language | English |
Abstract
Safety verification plays an essential role in analyzing whether a system works according to the specification requirements. Its objective is to verify if it is possible to reach the control objective without hitting a forbidden state - an obstacle. In a nutshell, safety is the problem of assigning the probability to an event. I will study the following configuration: a forbidden (unsafe) subset U of a state-space S and the set of initial conditions A. I aim at finding the largest probability that a process starts (deterministically) at a point of A and reaches U before it leaves the state space S. Based on the characterizations of safety, I will discuss several algorithms for computing safety.
In the second part of the talk, I will take the starting point of a Markov decision process (MDP) with a policy that, for each state, provides the probability of choosing a particular control action that is optimal and safe. The discussed solution to this challenge is to use the evolution equation, which relates the occupation measure with the hitting probability. The occupation measure corresponds to the expected number of the states' visits. Consequently, a safety function is derived from the evolution equation. Subsequently, we will present an online data-driven method to assess the safety of an MDP.
Bio
Rafal Wisniewski is a professor and a deputy head for research in the Department of Electronic Systems, Aalborg University.
He received his PhD in Electrical Engineering in 1997 and PhD in Mathematics in 2005. He was a control specialist at Danfoss A/S, later a Vestas professor. He spent his sabbaticals at NASA Goddard Center, MD, Mittag-Leffler Institute in Stockholm, Rise University, TX, and ETH, Zürich. His research interest is in system theory, particularly the theory of decisions and learning. For details of his research group please see https://www.learning-decision.dk
Safety verification plays an essential role in analyzing whether a system works according to the specification requirements. Its objective is to verify if it is possible to reach the control objective without hitting a forbidden state - an obstacle. In a nutshell, safety is the problem of assigning the probability to an event. I will study the following configuration: a forbidden (unsafe) subset U of a state-space S and the set of initial conditions A. I aim at finding the largest probability that a process starts (deterministically) at a point of A and reaches U before it leaves the state space S. Based on the characterizations of safety, I will discuss several algorithms for computing safety.
In the second part of the talk, I will take the starting point of a Markov decision process (MDP) with a policy that, for each state, provides the probability of choosing a particular control action that is optimal and safe. The discussed solution to this challenge is to use the evolution equation, which relates the occupation measure with the hitting probability. The occupation measure corresponds to the expected number of the states' visits. Consequently, a safety function is derived from the evolution equation. Subsequently, we will present an online data-driven method to assess the safety of an MDP.
Bio
Rafal Wisniewski is a professor and a deputy head for research in the Department of Electronic Systems, Aalborg University.
He received his PhD in Electrical Engineering in 1997 and PhD in Mathematics in 2005. He was a control specialist at Danfoss A/S, later a Vestas professor. He spent his sabbaticals at NASA Goddard Center, MD, Mittag-Leffler Institute in Stockholm, Rise University, TX, and ETH, Zürich. His research interest is in system theory, particularly the theory of decisions and learning. For details of his research group please see https://www.learning-decision.dk
Practical information
- General public
- Free
Organizer
- Professor Maryam Kamgarpour
Contact
- chantal.demont@epfl.ch barbara.schenkel@epfl.ch