Erik Sandewall: Features and Fluents, Volume I

Published by Oxford University Press, 1994.

Key Results

An ontological framework is defined for situations where an agent performs actions in some world. The framework is structured as a game between an ego and a world, and serves as an underlying semantics for the logics: it defines what is the set of intended models (and therefore, the set of intended conclusions) for a given scenario description.

The ontological framework is first used for defining an ontological and epistemological taxonomy for chronicles (that is, scenario descriptions). In this taxonomy, one can characterize classes of chronicles based on properties such as e.g. whether actions are deterministic; whether there are any observations for timepoints later than zero; whether actions are assumed to take only a single timestep each; and so on. These classfications are then used in the assessments of entailment methods.

Next, a specialized first-order logic is introduced for expressing scenario descriptions as logical formulae. The logic uses essentially a predicate of the form

   Holds(t,f,v)

where t is a timepoint, f is a feature symbol (a "fluent"), and v is the value of that feature at time t. Timepoints are interpreted as nonnegative integers. A somewhat different syntax is used for reasons of convenience, especially when writing chronicles of nontrivial size, so the same formula would actually be written

   [t] f = v

with a non-standard equality sign (with a ^ over the =).

As usual, the set of classical (Tarski) models for a given chronicle is a superset of the intended models, since the intended models require inertia (persistence) and the classical models don't. Assuming that the set of selected models can be chosen as those classical models which are minimial wrt. a particular preference relation on models, we define the range of applicability of a preference relation (or more generally, an entailment method) as the set of chronicles for which the set of selected models equals the set of intended models. (An entailment method is contructed by combining one or more preference relations with other operations, in particular using a partitioning of the set of premises, and using filtering).

About a dozen entailment methods are defined and analyzed in the book. Several of them have previously been proposed in the literature, but some are new ones or improvements on previous ones. The range of applicability is identified for each one of them.

In order to determine whether the range of applicability so obtained can be improved, we also address the question of the upper bounds of the range of applicability. The upper bound is, in principle, a set of chronicles such that for any chronicle outside the set, the set of selected models is different from the set of intended models. (The exact definition is slightly more complex, in order to be practically meaningful). In the book, upper bounds on the range of applicability are identified for all of the entailment methods that have been defined. In many cases the upper bound equals the range of applicability (which in itself is a lower bound, of course), meaning that the obtained range is exact and can not be improved.

The present volume of this work applies to chronicles which may have the following properties:

Non-deterministic actions.
Observations at any time (not only initial ones), which means that both prediction and postdiction are included.
Incompletely known timepoints for the starting and ending of actions, and for observations.
Actions with extended duration in time.
Metric (integer time) which may be either linear or branching. Branching metric time contains situation calculus as a special case.
Composite actions in the scenarios, formed using sequential composition, conditional actions, and repetition of actions.

However, the analysis is restricted to scenarios respecting the following assumptions and restrictions:

No qualification, ramification, or surprises.
Sequential execution of the actions - no concurrency.
No cause-and-effect chains.
Discrete time and discrete features (finite set of possible values) - no continuous change.
The agent has complete knowledge of the world - no account of information-gathering actions.

The current work in our group aims at extending the same systematic approach to broader classes of scenarios, where one or more of these restrictions have been lifted. Bibliographic and ordering information for this book are obtained by clicking here.