The distant goal of my research has always been to find ways to facilitate human involvement in constructing and accepting probabilistic AI systems in data-poor domains. My approach is mostly focussed on studying and exploiting (mathematical) properties of Probabilistic Graphical models (Bayesian networks) to design new methods for their construction and explanation, as well as on understanding the effects of various precision-complexity tradeoffs in the specification of such models.
Zie ook mijn onderzoekspagina of mijn pagina met projecten.
The aim of this project is to design simulation-based processes for creating and evaluating Bayesian Networks that can be applied to real-world problems, such as the problem of the reference class, problem of the priors, causality, and others - or to show limitations of such an approach. The process should be hybrid, in that it plays to the strengths of both the BN and human modellers and users (arguers) by combining probabilistic updating rules with rational deliberation to decide which events, causalities and probabilities to apply.
Ons doel is om systemen te ontwikkel die KI systemen kunnen monitoring, zodat er gecheckt wordt of het KI systeem zich houd aan voorgelegde beperkingen zoals normen of protocollen. Deze beperking zijn het resultaat van de omstandigheden waarin het KI systeem opereert en kunnen bijvoorbeeld het resultaat zijn van gelimiteerde middelen, wetten en regels en ethische of sociele overwegingen. Het monitor systeem moet deze voorgelegde beperkingen modelleren op een voor mensen intuitive en transparante manier, zodat de betrokken deze kunnen blijven bijwerken en het mogelijk is om uit te leggen aan welke beperkingen het KI systeem zich houdt en welke niet. Dit zorgt voor communicatie en samenwerking tussen de mensen en KI systemen.
By investigating the relations between knowledge, reasoning and data, we aim to develop mechanisms for the verification and evaluation of hybrid systems that combine manual knowledge-based design and learning from data. The focus will be on structures used in reasoning and decision-making, in particular logical and probabilistic relations (as in Bayesian networks) and reasons (pro and con) and exceptions (as in argument-based decision making).
In Responsible HI, there is a need to ensure that the behaviour of HI systems is aligned with legal and moral considerations. So the input and output of a system must meet a given set of principles. Verification, evaluation and interaction mechanisms are needed that ensure such alignment, that show to what extent alignment has been achieved, and that help improve the alignment.
Bayesian networks (BNs) provide decision support in complex investigative domains where uncertainty plays a role, such as medicine, forensics and risk assessment. Yet, BNs are only sparsely used in practice. In data-poor domains, they have to be manually constructed, which is too time-consuming to support pressing decisions. Furthermore, few domain experts have the mathematical background to build a BN, a graph representing dependencies among variables with probability distributions over these variables. So despite the increased analytical power a BN could bring with respect to, for example, evidence aggregation or sensitivity analysis, many experts still use more qualitative concepts such as scenarios (stories, cases, timelines) and arguments (evidence graphs, ordered lists), which convey verbally expressed uncertainty ("strong evidence", "plausible scenarios").
If BNs are to be used in actual investigations, we need software tools and interfaces for BN construction that are engineered into the heart of the decision-making process. These tools should be based on familiar, more linguistically-oriented concepts such as arguments and stories, and complemented by algorithms intended to speed up and facilitate the BN-building process.
Lucia de Berk heeft het ondervonden: bewijs op basis van statistiek leidt gemakkelijk tot fouten. Dit project beoogt het maken van zulke fouten te helpen voorkomen. De nieuwe aanpak van het project is om de succesvolle statistische modelleertechniek van Bayesiaanse netwerken te koppelen aan goed bij de juridische denkwereld aansluitende modellen van argumentatie en scenarioconstructie.
Dit project bestaat uit twee sub-projecten. In het ene project (aio: Sjoerd Timmer, UU) worden Bayesiaanse netwerken en argumentatie aan elkaar verbonden; in het andere project (aio: Charlotte Vlek, RUG) is de focus op het verbinden van Bayesiaanse netwerken met scenario-gebaseerde methoden. Voor beide aio's fungeer ik als co-promotor.