Belief Management for Autonomous Robots Using History-Based Diagnosis - Developing Concepts in Applied Intelligence - page 113

task, nor to continue operation from a sane situation. Thus, in order to ensure quality
of service, a reasoning mechanism and belief management is needed that is able to deal
with this ambiguity by considering multiple hypotheses.
For our implementation of such a belief management system, based on the agent pro-
gramming language IndiGolog, we use the situation calculus and adapt history-based
diagnosis [3] developed by Iwan and colleagues. Our system allows us to detect incon-
sistencies, describe belief ambiguities, and generate multiple hypotheses as well as rank
them. While we adopt the favorite hypothesis for immediate operation, we keep track
of the alternatives in case the chosen favorite is proven to be wrong by future data.
The remainder of this paper is organized as follows. Related work is covered in
Section 2. In Section 3 we depict our belief management system, where section 3.1
contains some preliminaries and Section 3.2 covers the details of our system. Experi-
ments can be found in Section 4, followed by conclusions drawn in Section 5.

2 Related Research

In the following we discuss previous work in three research areas that are relevant to
our paper: (1) fault diagnosis, (2) hypothesis discrimination and (3) dealing with uncer-
tainties in acting and sensing.
Diagnosis, i.e. the detection and localization of faults, is a major topic in research.
Many approaches like [4,5] specifically address the diagnosis of sensing and/or actuator
faults. For complex and dynamic environments, classic diagnosis like consistency-based
diagnosis [6] is unfortunately too static. While it blames faulty components, for dy-
namic environments it is more appropriate to focus on correct action (event) sequences
[7]. In [3], Iwan describes a diagnosis approach for robots in dynamic environments
that is based on the situation calculus, which we will discuss in Section 3.1.
Once a set of diagnoses has been found, the correct one has to be isolated, or at least
we have to identify the most probable one(s). In [8] the authors proposed a planning
method to derive action sequences that improve the diagnosis quality by ruling out in-
appropriate ones. Handling uncertainty in sensing and acting is another issue relevant
to our work. There are several approaches using the situation calculus to deal with this
problem. For example, formalizing a planning problem, one can decide the relevance
of unexpected changes for an actual plan [9]. The authors of [10] employ situation cal-
culus as well as decision-theoretic planning to control an autonomous robot. Whenever
passive sensing differs from expectations, re-planning is initiated. In [11] Thielscher
et al. presented a formalization using the fluent calculus to derive explanations for un-
expected situations. Their approach employs active sensing to minimize the impact of
such events. Weld et al. [12] extended GraphPlan to handle sensing actions and uncer-
tainty, but they do not maintain the consistency of a system's belief.

3 History-Based Diagnosis and Belief Management for IndiGolog

In this section we present a belief management system that ensures the consistency of
a robot's belief with reality (as it is perceivable), preceded by necessary preliminaries.