Abstract |
The fields of Artificial Intelligence (AI) and Robotics were strongly connected in the early days of
AI, but have since diverged as practitioners of AI focused on problems and algorithms abstracted
from the real world, while roboticists, building on their
background in mechanical and electrical
engineering, concentrated on sensory
-
motor functions. With advancements in both fields, there
is now a renewed interest in bringing the two disciplines closer together. Robots, or any other
autonomous entity inhabiti
ng real
-
world domains, need to deal with incomplete information and
uncertainty at various levels of abstraction, from low
-
level sensory data to high
-
level knowledge,
such as action preconditions and effects. This thesis concentrates on the latter, aiming
to keep
the Knowledge Base (KB) of an agent both up
-
to
-
date and consistent, while performing world
-
changing or observation actions.
Action theories are well
-
established logical theories, based on classical logic, for reasoning about
domains involving dyn
amically changing environments. Thus, they can inherently deal with
change caused by actions. One of the most prominent action languages is the Event Calculus (EC),
which incorporates certain useful features for representing causal and narrative informatio
n that
differentiate it from other similar formalisms. The EC explicitly represents temporal knowledge,
enabling reasoning about the effects of a narrative of events along a time line. Given that the
logical theory stored in a KB is not always correct, the
re is also a need to revise KBs as new
information is received. The area of belief revision addresses such a change to a KB. In the well
-
known AGM postulates, belief revision emerges when one has a knowled
ge base K and a formula
α
, and the issue is how to
consistently incorporate
α
in K to obtain a new KB K'. This means that
some of the beliefs in the original KB must be retracted, but not all of them, since this would be
an unnecessary loss of valuable information. What makes things more complicated is tha
t beliefs
in a knowledge base have logical consequences, so when giving up a belief one has to decide as
well which of the consequences to retain and which to retract. Thus, belief revision is non
-
trivial
as several different ways for performing this opera
tion may be possible. From the EC perspective,
there has been extensive work on epistemic extensions of action languages, in general, as well as,
on the main EC formalism. However, little attention has been paid to the problem of automatically
revising (co
rrecting) a KB in the EC when an observation contradicts the already inferred
knowledge, despite mature work on the belief revision field.
As the current trend in related research is to identify efficient ways to couple high
-
level task
planning with low
-
level task execution or feasibility checking, the current work aims to empower
such combinations, through the delivery of a more generic high
-
level formalism that lifts some of
the unrealistic assumptions of existing solutions. We propose a generic framew
ork in the context
of the EC, along with Answer Set Programming (ASP) encodings of the revision algorithm,
accommodating belief revision on top of EC axiomatizations. We consider both the epistemic and
non
-
epistemic case, relying on the possible
-
worlds rep
resentation to give formal semantics to an
agent's belief state. We formalize notions of commonsense revisions that take into consideration
different knowledge states, such as factual (or observed) knowledge, default, inferred and also
unknown knowledge. W
e present a methodology and an ASP encoding that can implement the
formalism, in which we adapt the existing powerful action theories in a more realistic setting. We
also present an optimization algorithm aiming to improve the efficiency of the implementat
ion.
Finally, we discuss possible future expansions and improvements of our framework and how this
work can form the substrate for further extensions concerning a richer set of commonsense
features, along with formal results showing that it is generic enou
gh to be applied to different EC
dialects.
|