Generation of Virtual Test Scenarios for Training and Validation of AI-based Systems

Dahmen, U.; Osterloh, Tobias; Rosmann, Jurgen

doi:10.1109/pic53636.2021.9687075

“…AI models are complex, opaque and challenging for humans to understand (van Nuenen et al, 2020), especially if the decision making cannot be observed and remains a black box (Käde & Maltzan, 2019). Even for developers it can be challenging to maintain oversight, amid the selflearning capabilities of algorithms when discovering new causal relationships (Walmsley, 2021), and parts of the system cannot be assessed independently (Dahmen et al, 2021). Hence, even algorithm designers cannot provide full transparency (Walmsley, 2021).…”

Section: Socio-technical Limitationsmentioning

confidence: 99%

“…Assessors struggle to guarantee that an AI-based system is free from bias and discrimination because over time existing stereotypes can be learned and reinforced (Käde & Maltzan, 2019;Walmsley, 2021). Even if excluded for a certain in time, due to impermanence it cannot be predicted whether new attributes will be learned that result in discrimination and bias (Dahmen et al, 2021), making it challenging to check that a system is not discriminatory at a given point in time to deduce future guarantees. An important research question is thus: How can AI impermanence be conceptualized and managed in a way that AI assessments can guarantee AIbased systems' long-term safety?…”

Section: Socio-technical Limitationsmentioning

confidence: 99%

Why it Remains Challenging to Assess Artificial Intelligence

Brecker,

Lins,

Sunyaev

2023

Proceedings of the Annual Hawaii International Conference on System Sciences

0

View full text Add to dashboard Cite

Artificial Intelligence (AI) assessment to mitigate risks arising from biased, unreliable, or regulatory noncompliant systems remains an open challenge for researchers, policymakers, and organizations across industries. Due to the scattered nature of research on AI across disciplines, there is a lack of overview on the challenges that need to be overcome to move AI assessment forward. In this study, we synthesize existing research on AI assessment applying a descriptive literature review. Our study reveals seven challenges along three main categories: ethical implications, regulatory gaps, and socio-technical limitations. This study contributes to a better understanding of the challenges in AI assessment so that AI researchers and practitioners can resolve these challenges to move AI assessment forward.

show abstract

“…The scenario configuration stage involves the definition of the basic application scenario. This paper uses the classification of Dahmen et al by classifying scenarios into abstract, logical and concrete scenarios [6].…”

Section: Scenario Configurationmentioning

confidence: 99%

Scenario-Driven Data Generation with Experimentable Digital Twins

Maqbool¹

2022

Preprint

0

View full text Add to dashboard Cite

Synthetic data is an indispensable supplement to the difficult-to-acquire real data in order to meet the substantial demand by machine learning based systems. Data playing the key role in machine learning models, its objective and maintainable quality metrics are vital for quality assurance of the whole system. This paper introduces a systematic and domain-neutral methodology based on formalized scenario variation and experimental digital twins for the generation of synthetic data. The methodology uses human-readable scenarios and semantically meaningful parameter variations to describe possible entities, actions and events to be simulated, whereas experimental digital twins bring the scenarios to life by the integration of various domains of a system such as mechanics, sensors, actuators and communication under one platform that can be simulated as a whole. The scenario description and digital twin simulation is carried out iteratively to derive the optimal distribution of synthetic data. Thus scenarios and experimentable digital twins can together serve as mediums to systematically cover diverse application scenarios, test dangerous situations and find faults within a system.

show abstract

Scenario-Driven Data Generation with Experimentable Digital Twins

Maqbool

¹

,

Roßmann

²

2023

Annals of Scientific Society for Assembly, Handling and Industrial Robotics 2022

0

View full text Add to dashboard Cite

Synthetic data is an indispensable supplement to the difficult-to-acquire real data in order to meet the substantial demand by machine learning based systems. Data playing the key role in machine learning models, its objective and maintainable quality metrics are vital for quality assurance of the whole system. This paper introduces a systematic and domain-neutral methodology based on formalized scenario variation and experimental digital twins for the generation of synthetic data. The methodology uses human-readable scenarios and semantically meaningful parameter variations to describe possible entities, actions and events to be simulated, whereas experimental digital twins bring the scenarios to life by the integration of various domains of a system such as mechanics, sensors, actuators and communication under one platform that can be simulated as a whole. The scenario description and digital twin simulation is carried out iteratively to derive the optimal distribution of synthetic data. Thus scenarios and experimentable digital twins can together serve as mediums to systematically cover diverse application scenarios, test dangerous situations and find faults within a system.

show abstract

Generation of Virtual Test Scenarios for Training and Validation of AI-based Systems

Cited by 8 publications

References 5 publications

Why it Remains Challenging to Assess Artificial Intelligence

Why it Remains Challenging to Assess Artificial Intelligence

Scenario-Driven Data Generation with Experimentable Digital Twins

Scenario-Driven Data Generation with Experimentable Digital Twins

Contact Info

Product

Resources

About