Hilke Garbe

Probabilistic and Empirical Grounded Modeling of Agents in (Partial) Cooperative Traffic Scenarios

The Human Centered Design (HCD) of Partial Autonomous Driver Assistance Systems (PADAS) requires Digital Human Models (DHMs) of human control strategies for simulations of traffic scenarios. The scenarios can be regarded as problem situations with one or more (partial) cooperative problem solvers. According to their roles models can be descriptive or normative . We present new model architectures and applications and discuss the suitability of dynamic Bayesian networks as control models of traffic agents: Bayesian Autonomous Driver (BAD) models. Descriptive BAD models can be used for simulating human agents in conventional traffic scenarios with Between-Vehicle-Cooperation (BVC) and in new scenarios with In-Vehicle-Cooperation (IVC). Normative BAD models representing error free behavior of ideal human drivers (e.g. driving instructors) may be used in these new IVC scenarios as a first Bayesian approximation or prototype of a PADAS.

KARaCAs: knowledge acquisition with repertory grids and formal concept analysis for dialog system construction

We describe a new knowledge acquisition tool that enabled us to develop a dialog system recommending software design patterns by asking critical questions. This assistance system is based on interviews with experts. For the interviews we adopted the repertory grid method and integrated formal concept analysis. The repertory grid method stimulates the generation of common and differentiating attributes for a given set of objects. Using formal concept analysis we can control the repertory grid procedure, minimize the required expert judgements and build an abstraction based hierarchy of design patterns, even from the judgements of different experts. Based on the acquired knowledge we semi-automatically generate a Bayesian Belief Network (BBN), that is used to conduct dialogs with users to suggest a suitable design pattern for their individual problem situation. Integrating these different methods into our knowledge acquisition tool KARaCAs enables us to support the entire knowledge acquisition and engineering process. We used KARaCAs with three design pattern experts and derived approximately 130 attributes for 23 design patterns. Using formal concept analysis we merged the three lattices and condensed them to approximately 80 common attributes.

Predicting the focus of attention and deficits in situation awareness with a modular hierarchical Bayesian driver model

Situation Awareness (SA) is defined as the perception of elements in the environment within a volume of time and space, the comprehension of their meaning, and the projection of their status in the near future [1]. Lacking SA or having inadequate SA has been identified as one of the primary factors in accidents attributed to human error [2]. In this paper we present a probabilistic machine-learning-based approach for the real-time prediction of the focus of attention and deficits of SA using a Bayesian driver model as a driving monitor. This Bayesian driving monitor generates expectations conditional on the actions of the driver which are treated as evidence in the Bayesian driver model.

A greedy knowledge acquisition method for the rapid prototyping of knowledge structures

The main goal of this paper is the presentation of a new GReedy knowledge Acquisition Procedure (GRAP) for rapid prototyping of knowledge structures (KS) or spaces. The classical knowledge acquisition method for this [2] is even for domain experts cognitive demanding and computational complex. GRAP interactively generates an online knowledge acquisition schedule so that experts only have to provide simple nonredundant judgements about the (learning / cognitive) precedence in pairs of (learning / cognitive) objects. From these data GRAP generates a Hasse diagram of the surmise relation from which the knowledge structures and optimal user-adaptive learning paths can be derived. In a case-study we developed with three expert software engineers a knowledge structure and optimal learning paths for 23 software design patterns within a few hours.

Learning the DAG of bayesian belief networks by asking (conditional) (in-)dependence questions

Bayesian belief networks (BBNs) have become the de facto standard for the representation of uncertain knowledge. They consist of a qualitative and of a quantitative part describing the (in-)dependencies between the variables of interest as a directed acyclic graph (DAG) and the decomposition of the joint probability distribution (JPD) as a product of conditional probability distributions constrained by the structure of the DAG. In this paper we present a new constraint-based query procedure: Query-an-Oracle (QAO). We assume that an oracle -- preferable a human domain expert -- is at hand which is competent and willing to answer questions generated by QAO concerning the directed (causal) dependence and (conditional) independence of the relevant random variables in the domain. Compared to other structure learning methods (e.g. the PC-Algorithm of Peter Spirtes and Clark Glymour and the IC-Algorithm of Pearl) QAO has a number of advantages. It derives the DAG of the BBN with less computational complexity, with no redundant questions, and is able to exploit directed dependence information without urging oracles to differentiate between direct and indirect influence.