John-Jules Ch. Meyer and Jan Treur
A currently popular definition of artificial
intelligence (AI) is: ‘the study of agents that exist in an environment
and perceive and act’. Agents, often referred to as intelligent agents,
are (hardware or software) entities that act on the basis of a ‘mental state’.
They possess both informational and motivational attitudes, which means that
while performing their actions they are guided by their knowledge and beliefs
as well as their desires, intentions and goals, and,
moreover, they are able to modify their knowledge, intentions, etc. in the
process of acting as well.
Clearly the description of agent behaviour involves
reasoning
about the dynamics of acting, and if agents are supposed
to be deliberative, they should also themselves be able to do so. Furthermore,
since the actions of agents may - apart from actions that change the external
world directly - also include reasoning as a special kind of mental action (for
example, performing some belief-revising action or an action comprising of
reasoning by default), it may be clear that in the context of agent systems the
dynamics of reasoning and reasoning about dynamics go hand in hand.
One of the recognized problems in AI is the gap
between applications and formal foundations. This sixth volume, as well as the
following one in the DRUMS series, does not present the final solution to this
problem, but at least does an attempt to reduce the gap by bringing together
state-of-the-art material from both sides and by clarifying their mutual
relation. One of the areas where the application-foundation gap is deeply felt
is the area of reasoning processes as they occur within agents.
In practical applications complex (expert) reasoning
processes are modelled, both in the context of knowledge-based systems for
complex tasks and in the context of intelligent agents. In such practical
reasoning processes it is often important, given the aims of the reasoning, to
minimize the amount of information needed, and thus make (rational) decisions
about the focuses of the reasoning process. Therefore models of such practical
reasoning processes, have to address complex dynamics, including automated
solutions for
For example, in models for diagnostic reasoning
processes almost all of these aspects of dynamics often occur. But also for
information gathering agents these dynamic aspects play an important role. From
the side of foundations of reasoning processes, since 1980 a lot of work has
been done, in particular on logics for nonmonotonic reasoning. However, very
few contributions have addressed the dynamics of these reasoning processes, and
no contribution at all has covered all aspects listed above.
More recently, the area of agent models has
been addressed. For example, in order to formalize agents with belief, desires
and intentions (BDI-agents), specific logics have been developed. However, the
architectures for BDI-agents that have been developed for practical
applications are not related in a clear manner to these logics. In particular,
although some degree of dynamics is captured by them, BDI-logics lack a clear
relationship to the internal dynamics of the reasoning in applications of
BDI-agents: e.g., the revision of beliefs, desires and intentions under the
appropriate circumstances, have not been sufficiently addressed in the
foundational approaches. In applications of BDI-agents, this rather complex
issue of revision of beliefs, desires and/or intentions has to be addressed,
and actually has been addressed; e.g., when to revise only an intention and not
the underlying desire, when to revise both, when to revise a desire and still
keep a (committed) intention which was based on the desire?
In a comparable manner, existing theories of diagnosis
from first principles address diagnosis from a static perspective, but do not cover
process-oriented questions such as, on which hypothesis should the process be
focused, and which observations should be done at which moment. Similarly, most
contributions in the area of logics for nonmonotonic reasoning only offer a
formal description of the possible conclusions of such a reasoning process, and
not of the dynamics of the reasoning process itself.
In conclusion, as far as the relation to applications
is concerned, within these foundational approaches especially the dynamic
aspects of the agent's internal (reasoning)processes are often not, or at least
not fully covered. The areas of models for reasoning patterns and agents, show
a gap between applications and foundations in which the lack of adequate
dynamics incorporated in the foundations plays an important role. For this
reason in this volume the main theme is dynamics of reasoning processes. In the
next volume, also dynamics of the external world is addressed. Agents often
reason about both types of dynamics.
In this book we will consider various formal means
regarding the topics reasoning and dynamics, mostly from an intelligent agent
perspective. By formal means we mean formal logics / calculi as well as formal
specification languages together with their formal semantics. We will see how
in particular temporal and dynamic logic / semantics are useful means to
perform this formal analysis. (Since these frameworks play such a prominent
role in this book, we have included a separate ‘Basic Concepts’ chapter following
this introduction, where these are explained succinctly.)
Formal foundations of reasoning models are of
importance for different reasons. First, by defining semantics in a formal
manner, a precise and unambiguous meaning of the syntactical constructs is
obtained, which may help designers. This requires that designers are familiar
with the formal techniques used to define such formal semantics. Unfortunately,
this requirement is often not fulfilled for application developers in practice,
and there is no reason to expect that this will change in the short term.
However, more realistically, those who develop a modelling technique often have
more knowledge of formal methods. Therefore they can benefit a lot from
knowledge of formal foundations during development of their modelling
technique, and use that also as a basis to informally or semi-formally describe
the semantics for others (application developers using the modelling technique)
with a less formal background.
Secondly, formal foundations are especially important
to obtain the possibility of verification of a design or verification of
requirements. Verification is usually a rather technical and tedious matter,
only feasible for specialists (‘verification engineers’). They need to know
about the formal foundations, including formal semantics and proof systems.
In this volume the emphasis is on the investigation of
the dynamics of reasoning processes within an agent. Reasoning takes (place in)
time, so, for example, it is natural to view the behaviour of reasoning
processes from a temporal perspective, and consider a temporal semantics of
these processes. For example, in meta-level architectures the reasoning may
switch from object level to meta-level, and one may describe such behaviour by
means of temporal models. Furthermore, to reason about this temporal behaviour
it is very natural to employ temporal logic of some kind. But, of course,
taking a temporal stance is possible for the analysis of any reasoning process
or system, like (multi-)agent systems.
A large variety of reasoning patterns is presented in
this volume: from dynamic generation and retraction of assumptions and dynamic
control of reasoning in meta-level architectures to diagnostic reasoning
processes and non-monotonic reasoning processes. All these reasoning processes
have in common that they are defeasible in some sense, that is, it is
possible (or perhaps even typical) that at a certain stage of the reasoning
process certain conclusions may be (tentatively) arrived at, which possibly
should be abandoned at a later stage. This type of reasoning is generally
called defeasible reasoning. In the first two papers meta-level
reasoning is exploited to focus the object level reasoning process, either by
dynamically focusing on a set of goals for the object level reasoning, or on a
set of assumptions used as premises in the object level reasoning. The other
papers in this book address the dynamics of different variants of nonmonotonic
reasoning such as reasoning by default. Here ‘nonmonotonic reasoning’ means a
form of (nonstandard) logical reasoning in which adding assumptions to a set of
premises may have loss of derivable conclusions as a result. This is, of
course, nonstandard in the sense that classical first-order logic is
‘monotonic’ (not nonmonotonic). On the other hand a common sense reasoning
method like reasoning by default is typically nonmonotonic: conclusions may be
drawn on the basis of a set of premises, inclusing tentative ones on the basis
of lack of information, while adding information to the premises may result in
not having certain conclusions derivable any more. Clearly, a form of
nonmonotonic reasoning in the above sense is also defeasible: if one considers
the reasoning process in time clearly it may be the case that as more
information comes available as time progresses, tentative conclusions (e.g.,
default conclusions) have to be retracted. Due to the emphasis of DRUMS on
nonmonotonic reasoning this type of reasoning patterns is discussed quite extensively.
In the first paper of this book temporal semantics is
defined for a goal-directed reasoning process in which explicit reasoning about
the goals of reasoning takes place, using a meta-level architecture. By this
approach it is possible to dynamically determine the goals of the reasoning by
meta-reasoning. The semantics makes use of combined linear time temporal models
for object level and meta-level.
The second paper addresses the dynamics of reasoning
processes in which assumptions can be added and retracted dynamically, on the
basis of meta-level reasoning. One of the applications of this approach is
reasoning by indirect proof. Another application, addressed in the paper, is
diagnosis based on causal knowledge. The semantics is based on branching time
temporal models over epistemic states.
The third paper addresses the dynamics of reasoning in
a generic model for a diagnostic task, which has been applied, among others, to
chemical Nylon production processes. It is shown how a temporal semantics approach
to these diagnostic reasoning processes can be used to verify dynamic
properties of the reasoning process.
The following nine papers address the dynamics of
nonmonotonic reasoning patterns such as default reasoning.
The fourth paper presents a general (temporal)
semantic framework for nonmonotonic reasoning processes. Within this framework
it is possible to specify the dynamics of a large class of nonmonotonic
reasoning patterns.
The fifth, sixth and seventh paper address the
dynamics of nonmonotonic reasoning based on default logic.
The fifth paper defines a temporal semantics for
default reasoning processes. An interpretation mapping of Reiter's default
logic into linear time temporal epistemic logic is defined. According to this
interpretation, the antecedent of a default rule refers to the past, the
justification refers to the future of the reasoning process, and the consequent
to the current time point.
In the sixth paper, the KARO framework (here to be
considered as a form of doxastic dynamic logic) is used to formalise (a
particular type of) default reasoning processes. The dynamics of reasoning by
default is considered from the viewpoint of an agent which performs default
reasoning steps as particular actions, and is thus analyzed by using the
dynamic logic operators together with the doxastic ones in the KARO framework.
In the seventh paper another dynamic logic approach is
used to describe the dynamics of default reasoning processes. In this approach
also a blend of epistemic/doxastic and dynamic logic operators is employed. The
treatment of default reasoning is more general than in the previous approach.
(In some sense it is reminding of the use of dynamic logic for describing
reasoning steps in more general reasoning systems by Sierra et al. that
we will encounter in Volume 7.) On the other hand the actions are confined to
the application of default rules, which is thus not integrated into a general
framework of actions performed by agents as in the previous approach.
An operational problem with default logic is that in
the reasoning, also the context of the future of the reasoning processes has to
be taken into account. Therefore, other, more constructive variants have been
proposed. Two of them, Temporal Epistemic Default Logic (TEDL), and
Constructive Default Logic (CDL) are presented in paper eight and nine.
The eighth paper addresses a variant of default logic
which is defined in a temporal context. The logic TEDL is introduced in which
default reasoning steps can be specified based on a branching time temporal
logic.
The ninth paper presents Constructive Default Logic
and specifies the control of default reasoning processes based on this logic.
It exploits selection knowledge to specify conflict resolution in case of
conflicting defaults.
The tenth paper addresses a situation calculus
approach to the dynamics of logic program execution.
In the eleventh paper a natural-deduction-based
approach to non-monotonic reasoning is presented. In this approach aspects of
the dynamics of the (nonmonotonic) reasoning process are incorporated into the
logic itself by means of so-called ‘contexts’, which, at any stage of the
deduction, contain the premises used sofar. Contexts thus are updated during
the reasoning process, and (nonmonotonic) conclusions must always be viewed
with respect to the context at hand.
In the twelfth paper in this part we encounter a
slightly different approach of dealing with the dynamics of a reasoning
process: instead of using an explicit temporal logic Veltman-style update
semantics is used to model the reasoning process (here applied to normative or
deontic reasoning). The basic idea behind update semantics is that the
processing of assertions in a (here deontic) logic causes an update of the
(deontic) state of the processing agent, and as such this approach is another
good example of analyzing the dynamics of reasoning.
The thirteenth paper addresses an application, viz.
that to cooperative information gathering agents. In this paper it is shown how
properties such as succesfulness of the cooperation depend on dynamic
properties (such as reactiveness and pro-activeness) of the agents
participating in the cooperation. The agents reason about the control of
dynamics of a number of aspects of the cooperative process of information
gathering. In particular, reasoning is performed about when observations are to
be performed, when available information has to be communicated to other
agents, when requests have to be communicated, and when conclusions have to be
drawn on the basis of acquired information. Although the information states and
reasoning processes of the agents are dynamic, the actual world is assumed to
be not dynamic in this applications.
The fourteenth paper shows a model for the internal
dynamics of an agentbreasoning and acting on the basis of beliefs, desires and
intentions. This type of agent is able to reason about dynamics of the world,
but also about the dynamics of its own reasoning processes, for example, about
when to retract a desire, or an intention. Within this model (a formal
specification of a design), in particular, it is specified how the agent should
revise in an appropriate manner not only beliefs, but also desires, intentions
and commitments, on the basis of specific characteristics such as blind
commitment, single-mindedness or open-mindedness. Semantics of the model are
inherited of the generic semantics of the formal specification language in
which it is specified. However, the question how specific, dedicated semantics
for this type of reasoning processes can be obtained is left open, as a
challenge for further research. Only very recently some preliminary and partial
results on this issue are found in the literature.
The fifteenth paper addresses a model for deliberative
evolution in an agent society. An agent in this society is able to
deliberatively generate the goal to create a new agent with certain desired
(behavioural) properties. Moreover, the agent can deliberatively design a new
agent architecture that satisfies these properties, and after this deliberation
process is finished, it is able to plan and perform a creation action in the
material world by which the designed agent actually is created and starts
functioning in the society. In the first place reasoning about desires, goals,
intentions and actions takes place. Within this reasoning process, a
particularly complex pattern occurs: the design of an agent model that
satisfies the requirements. In the second place, to manufacture an agent based
on the generated design, reasoning about dynamics of the world is involved.
Here reasoning about the dynamics of the world is related to the creation of
new entities that perform themselves reasoning processes. As in the previous
chapter, also in this case semantics of the model are inherited of the generic
semantics of the formal specification language in which it is specified. Also
here, the question how specific, dedicated semantics for this type of
integrated reasoning and acting processes can be obtained is left open, as a
challenge for further research.