Probabilistic Modeling
What Motivates Us
The world is an uncertain place. To understand complex behaviors in the real world, such as the behavior of human drivers, manufacturing processes, or complex physical reactions, we need to explicity model and account for these degrees of uncertainty.
Our Approach
We are developing new methods in probabilistic modeling that address these challenges. Specifically, we are building statistical models that combine data-driven learning with prior engineering knowledge. Our approach allows us to accurately model prediction uncertainties originated in real-world scenarios, such as the behavior and interaction of vehicles in the real world.
Application
The AI techniques developed as part of this work are of substantial value to Bosch business units that need to accurately and correctly model, quantify and verify the performance of systems operating in the real-world.
Design of Dynamic Experiments
Having informative data is crucial for many AI applications. In the research field Design of Dynamic Experiments, we are working on methods for data generation and collection on physical systems such as test-bench as well as in virtual settings such as simulation.
Here, we equip model-based Design-of-Experiment (DoE) with AI-features, while conducting cutting-edge research in probabilistic data generation for dynamical systems driven by real-world applications.
Probabilistic data generation enables us to effectively simulate, validate and test real systems and, thus, accelerating the overall engineering development process.
At Bosch, we are dealing with dynamical experiments in different contexts, starting with individual sensors for temperature, moisture or gas composition up to highly complex systems behaviors like battery state-of-health.
Use Case
Effective Data Generation for Dynamics Modeling
Introduction
Imagine you could automate the process of data collection and model training for all your virtualized sensor projects!
At Bosch, we are dealing with dynamical systems on all scales and in all contexts, starting with individual sensors for temperature, moisture or gas composition up to highly complex systems behaviors like battery state-of-health or exhaust gas treatment systems in combustion vehicles. We are building tools that allow efficient and reliable modeling of these systems including tools to automate data collection, model selection and training. These tools allow to improve the performance of existing sensors, to replace some sensors by pure software functionality and to evaluate complex system behavior fully simulation driven.
Our Research
Data Generation for Hybrid Models
Having informative data is crucial for many AI applications such as modeling. We equip model-based DoE with AI-features for the collection of maximally informative data and develop hybrid modeling approaches using such data to enhance available physical models.
Iterative Model Learning Process
We develop a novel iterative and safe data collection method appropriate for the execution on dynamical systems and employ probabilistic techniques for iterative modeling and learning dynamical behavior, while exploiting physical domain knowledge.
Safe Interaction with Real Systems
When collecting the data on real systems, e.g., test-bench, critical regions of these systems are not known beforehand. It is therefore challenging to explore the dynamics while considering potentially critical regions. Our approach ensures safety for the systems when exploring the dynamics for data collection.
Automatic Execution of Test-Benches
Our approach reduces cost and effort for DoE planning and dynamics modeling. Additionally, our DoE approach can be employed to enable an automatic safe execution of test-benches.
-
For automatic execution of test-benches, e.g., during data generation for modeling of battery‘s behavior, Active Learning approaches are used to iteratively measure the system in real-time.
References
Alec Yu, H. S., Yao, D., Zimmer, C., Toussaint, M., & Nguyen-Tuong, D. (2021). Active Learning in Gaussian Process State Space Model. ECML/PKKD. [PDF]
Li, C-Y., Rakitsch, B., & Zimmer, C. (2022). Safe Active Learning for Multi-Output Gaussian Processes. AIStats. [PDF]
Schreiter, J., Nguyen-Tuong, D., Eberts, M., Bischoff, B., Markert, H., & Toussaint, M. (2015). Safe Exploration for Active Learning with Gaussian Processes. ECML/PKKD. [PDF]
Zimmer, C., Meister, M., & Nguyen-Tuong, D. (2018). Safe Active Learning for Time-Series Modeling with Gaussian Processes. NeurIPS. [PDF]
Probabilistic Modeling for Dynamical Systems
Modeling dynamical systems, while taking in account real-world uncertainty, is fundamental for many applications, such as root cause identification of validation failures.
In the focus area Probabilistic Modeling for Dynamical Systems, we build complex systems of dynamical prediction models while incorporating prior system knowledge derived from domain expertise.
Here, we are developing new methods that address technical challenges, such as efficient probabilistic inference, imperfect measurements and uncertainty quantification.
These approaches enable accurate and uncertainty-aware modeling of complex dynamics and validation of dynamical systems. Our research is beneficial for different domains at Bosch, such as the automotive field where accurate models of vehicles under real-world usage are required.
Use Case
Vehicle Interaction and Behavior Estimation
-
For autonomous vehicles, it is crucial to understand the environment by estimating the interaction between vehicles and their driving behavior.
Introduction
Imagine you could drive a car that reacts naturally to traffic participants by considering their imminent behavior!
Together with Bosch business units, BCAI is developing data-based probabilistic forecasting models of traffic participant behavior to enable next-generation driver assistance and automated driving functions. Today’s driver assistance functions rely on complex rule books and decision trees that have to be manually assembled. With the increasing resolution and, thus, data input of underlying sensors and the growing complexity of the desired assistance functionality, this manual approach becomes prohibitively expensive. We work on lightweight machine learning solutions that integrate seamlessly in the existing software stack to improve the overall system performance and prepare the path for future, higher-level, functionality. Our goal is to accurately simulate and predict the behavior of multi-agent systems in realistic driving situations. As this is an inherently stochastic domain, such behavior predictions need to come in terms of interactive probabilistic forecasts, while able to understand multiple scenarios that may result in one’s own vehicle’s actions.
Our Research
Accurate Behavior Prediction
Modern autonomous driving in particular driver assistance functions require accurate prediction of neighboring vehicles for a smooth and natural control of the ego vehicle. We develop AI-based approaches enabling a safe interaction in complex traffic situations.
Understanding Traffic Dynamics
Understanding the dynamic interactions between vehicles is highly challenging given that driving behavior is extremely complex and inherently stochastic. Our approach can cope with uncertainties in the interaction dynamics and provide accurate predictions for certain time horizon.
Multi-Agent Interaction Modeling
We develop novel methods using probabilistic graphical models and probabilistic neural networks for multi-agent system prediction. Our methods explicitly allow for modeling probabilistic systems while taking in account arbitrary numbers of agents and corresponding uncertainty quantification.
Real-World Application
In collaboration with Bosch business units, we develop and integrate AI-based models for accurate prediction of multiple traffic participants in real-world, e.g., for lane changes, while establishing a basis for future autonomous driving / driver Assistance functions.
-
Using multi-agent interaction modeling, driving decisions, e.g., lane changes, can be made for autonomous vehicles.
References
Herman, M., Wagner, J., Prabhakaran, V., Möser, N., Ziesche, H., Ahmed, W., Bürkle, L., Kloppenburg, E., & Gläser, C. (2021). Pedestrian Behavior Prediction for Automated Driving: Requirements, Metrics, and Relevant Features. IEEE Transactions on Intelligent Transportation Systems. [PDF]
Rudenko, A., Palmieri, L., Herman, M., Kitani, K. M., Gavrila, D. M., & Arras, K. O. (2020). Human Motion Trajectory Prediction: A Survey. The International Journal of Robotics Research. [PDF]