Abstract.
This paper
analyzes the meaning of intelligence within the domain of intelligent control.
Because of historical reasons, there are varying and diverse definitions of
what constitutes an intelligent control system. Due to the multidisciplinary
nature of intelligent control, there are no good formal measures for what is an
intelligent controller or for its performance. We argue that there are
characteristics which can be used to categorize a controller as being
intelligent. Furthermore, the measure of how intelligent a system is should
take into account a holistic view of the system’s operations and design
requirements. Beginning with the definition of intelligence characteristics and
performance measures, we can move towards a science of intelligent systems for
control.
Key words. Intelligent control, intelligent
systems, performance measures
1. INTRODUCTION
“Intelligent
Control” is a commonly used term, applied to an amorphous set of tools,
techniques, and approaches to implementing control systems that have
capabilities beyond those attainable through classical control. There is a
well-developed and deep body of work entailing a theory of control. This is not
the case for intelligent control, which is more of an ad hoc approach to
constructing systems that generally contain some internal kernel of the more
classical control, for which there is a theory.
We believe
that the discipline of intelligent control and its practitioners will benefit
from a more formal definition of what constitutes intelligence, as applied to
controllers. First of all, it will provide a framework within which the
researcher’s work can be understood. The understanding of each other’s work can
facilitate better progress in the field overall. Performance measures for
intelligent controllers can begin to be developed, providing benchmarks for
system performance and helping users and designers of systems better understand
the benefits of one approach versus another. The benefits would be based on
quantitative measures appropriate to the system requirements.
In this
paper, we analyze the history of intelligent control, its multidisciplinarity,
and the higher level issues of “meta control” that characterize intelligent
systems. We argue that there are characteristics that can be found in common
among most, if not all, systems labeled as intelligent controllers. These
characteristics suggest espousing a broader view of the elements of
intelligence that are necessary for advanced systems. Within this view, we
describe some possible levels of intelligence, which can guide the development
of and performance measures for controllers.
2. WHAT IS INTELLIGENT CONTROL?
In the late
seventies, Fu and Saridis introduced the combination of the words Intelligent Control. In the eighties, it
was brought to widespread practice and better understanding, Saridis, Meystel,
and Albus, who jointly spearheaded this movement.
As far as
methodology is concerned, the area was then and is now far from being stable.
The main obstacle to stability is due to the multidisciplinarity of the
intelligent control field. At the first Symposium on Intelligent Control in
1985, it was proclaimed a theoretical domain, in which control theory, AI, and
operations research intersected (Fig. 1 from).
Problems
arose due to the reluctance of members from the different constituent
disciplines to venture too far into the others’ domains. For instance,
specialists in control theory and automatic control tried to avoid getting
involved in issues of cognitive science, psychology, biology, and ecology
associated with the notion of intelligence. Specialists in the Artificial
Intelligence domain (itself comprised of various, rather diverse,
sub-specialties), traditionally abstained from getting involved in the
mathematical control issues related to system dynamics and felt uncomfortable
with domains of research that were beyond their discipline. The essence of the
conflict lies in the term “intelligent,” which has myriad implications. It is
still impossible today to distinguish an intelligent controller from one that
is not intelligent. It is also unclear whether a specialist in control theory
can be considered a specialist in intelligent control, or vice versa.
3.
THE TERM “INTELLIGENT CONTROL” AND ITS USAGE
As discussed above, the
problem of Intelligent Control has turned out to be an intrinsically
interdisciplinary one. This is why, starting in 1995, the IEEE and the National
Institute of Standards and Technology (NIST), in collaboration with other
agencies, organized a series of conferences intended to expand the topic of
Intelligent Control towards the more consistent, yet more difficult theme of
Intelligent Systems.
A recapitulation of the evolution of the term
Intelligent Control would be beyond the scope of this paper. We will present
some example of definitions and milestones in its development.
Fu linked a concept of
intelligent control with the following features that were traditionally out of
the scope of specialists in conventional control theory: decision making, image
recognition, adaptation to the uncertain media, self organization, planning,
and more .
Saridis gave the
definition of Intelligent Control as a statement of expected functions:
“Intelligent Control…would replace a human mind in making decisions, planning
control strategies, and learning new functions by training and performing other
intelligence functions whenever the environment doesn’t allow or doesn’t
justify the presence of a human operator…..Such systems… can solve problems,
identify objects, or plan a strategy for a complicated function of a system …
with intelligent functions such as simultaneous utilization of a memory,
learning, or multilevel decision making in response to fuzzy or qualitative
comments…”
In these excerpts, the
concept of goal is not mentioned, because goal is a part of a more general
statement, which includes intelligent control by necessity: “control of a
process employs driving the process to effectively attain a pre-specified goal”
.
A popular and
all-encompassing definition is by Albus: “… intelligence will be defined as an
ability of a system to act appropriately in an uncertain environment, where
appropriate action is that which increases the probability of success, and
success is the achievement of behavioral
sub goals that support the system’s ultimate goal.” Pang describes an
intelligent controller as a controller that is utilized for shaping the
behavior of an intelligent system . The distinct properties of this controller
are to provide the following features:
Ø It should
“know” what actions to take and when to perform them
Ø It should
reconcile the desirable and feasible actions
Ø It should
vary the high resolution details of control heuristics
Ø The
acquired control heuristics should be the most suitable ones and they should
change dynamically
Ø It should
be capable of integrating multiple control heuristics
Ø It should
dynamically plan the strategic sequence of actions
Ø It should
be able to reason between domain and control actions. In other words, it should
be able to use at least two levels of resolution simultaneously: the level of
the domain actions and the domain of the control actions.
In
their foreword to White and Sofge wrote, “To us, ‘intelligent control’ should
involve both intelligence and control theory. It should be based on a serious
attempt to understand and replicate the phenomena that we have always called
‘intelligence’ – i.e., the generalized, flexible and adaptive kinds of
capability that we see in the human brain.”
Cai
states that intelligent control possesses four features. These are
Ø It is a
hybrid control process, containing knowledge of mathematical and
nonmathematical models; it has no known single algorithm for dealing with the
complexity, incompleteness, and ambiguity it is confronted with.
Ø Its core is
in the higher level that organizes the problem solving. This higher level of
analysis and decision-making and planning requires related technologies, such
as symbolic information processing, heuristic programming, knowledge
representation, fuzzy logic, and automation of the discovery of similarity
amongst the solving processes. “There exists intelligence in the artificial problem-
solving process.”
Ø It is an
interdisciplinary field, requiring coordination with, and assistance from,
related fields, such as artificial intelligence, cybernetics, systems theory,
operations research, etc.
Ø It is a
research area under development only recently and would experience faster and
better development should a better theory for intelligent control be found.
4.
CHARACTERISTICS OF INTELLIGENT CONTROL
Some definitions of
intelligent control would limit it to those systems that rely on soft computing
techniques, such as fuzzy logic, neural networks, and genetic algorithms.
Another perspective is to analyze intelligent control systems with respect to
their characteristics, which were inherent in the quotes of the previous
section.
We begin by looking at
control laws, which are the heart of a control system, whether it is
intelligent or not. An analysis of control laws indicates that they are
subservient to the control system’s goal of reducing the deviation from a
pre-specified trajectory and/or the final state. It is more typical in the
present paradigm of automatic control to consider control laws to be part of a
broader “control strategy” for the overall system. The assignment for the
control law – to keep a certain variable of interest within some bounds around
a reference trajectory – comes as the result of some external intelligence
operating beyond the limits of the intelligence of the particular control law.
Intelligent controllers tend to incorporate these assignments (and their genesis)
into the function of the controller. This difference is a fundamental one.
In the fuzzy logic
controllers, the fundamental transformations move the set of input information
from the high resolution domain into the low resolution domain by using the tool
of “fuzzification.” Fuzzification plays the role of a generalization procedure,
because it searches for adjacency among information units, focuses attention,
and groups. Fuzzy logic controllers contain control mappings only for
generalized information at the lower resolution. When the required control
action is found, this lower resolution recommendation is instantiated or moved
to the domain of high resolution by the process of “defuzzification.” Hence,
fuzzy control systems are multi-resolution, and they move between resolutions
for the purposes of achieving the required control performance.
Neural networks are a
computation device for generalizing in vicinity (spatial or temporal). They are
a natural tool for moving information from higher to lower levels of
resolution. Therefore, neural networks are multiresolutional systems.
Expert System based
controllers make a judgment concerning generalizations and rules to be applied
at the lower resolution. They too presume a multi-level, multi-resolution
framework within which they operate.
Hybrid logic control
systems are multi-level control systems in which lower levels of resolution are
formulated in terms of logic based controllers, while the higher resolution
levels are analytical controllers (e.g., PID, Kalman). Again, there is at
minimum dual-level architecture of control, with differing resolutions.
Behavior-based controllers employ a concept of superposition of the activities
of multiple controllers working simultaneously, each providing for a separate type
of behavior. A hierarchical structure of some sort generally underlies
implementation. A low resolution level may select which behavior controller is
invoked at which time, or it may arbitrate between the different proposed
behaviors. Each individual behavior generation controller generates sub-goals
for itself. Although the previous examples are not an exhaustive analysis of
techniques and approaches for intelligent controllers, they serve to illustrate
that certain generalizations can be made about the essence of what
distinguishes a controller that is intelligent from one that isn’t. A Meta
level of control, which functions a lower level of resolution is necessary to
guide the underlying control system and expand its envelope of functioning.
5.
BEYOND INTELLIGENT CONTROL TO INTELLIGENT SYSTEMS
Saridis has defined the
intelligent control problem as “intelligent control is postulated as the
mathematical problem of finding the right sequence of internal decisions and
controls for a system structured according to the principle of increasing
intelligence with decreasing precision such that it minimizes its total
entropy”.
If intelligent control is responsible for
finding the right decisions and controlling the system to enact those
decisions, then there must be complementary supporting functions which provide
models of the world that the system is functioning in. There must be perception
and world modeling processes that occur alongside with the control function.
“Machine intelligence is the process of analyzing, organizing, and converting
data into knowledge, where (machine) knowledge is defined to be the structured
information acquired and applied to remove ignorance and uncertainty about a
specific task pertaining to the intelligent machine”. A procedural
characterization of an intelligent system is given by “intelligence is the
property of the system that emerges when the procedures of focusing attention,
combinatorial search, and generalization are applied to the input information
in order to produce the output” .
Key issues of the
domain highlighted include
Ø “A
desirable property of intelligent systems is that they are ‘adaptive’…
Ø Intelligence
is an internal property of the system, not a behavior…
Ø A
pragmatic reason for focusing on ‘intelligent’ control systems is that they
endow the controlled system with enhanced autonomy…”
Therefore, the overall system’s behavior
is what matters most in defining whether it is intelligent or not.
6.
HOW TO FORMALIZE INTELLIGENCE IN CONTROL
Taking a holistic view of
an “intelligent system,” we can begin to consider the formalization of the
organization of the system and how its performance is measured. One approach to
formalism is via an Elementary Loop of Functioning (ELF), which is common to
most concepts of intelligent systems. Fig. 2 shows the components of the ELF.
ELFs describe a closed loop of functioning common to intelligent systems. A
detailed description of the formation of hierarchies of nested loops is
described in .
It is our contention
that the emphasis in intelligent control has been placed primarily on the
Behavior Generation elements (and its constituents, such as planners and
executors). Broadening the scope to consider all the elements of ELF working
together within a hierarchy, we can consider the intelligence of the system as
a whole. The construction of a system based on the ELF model is driven by a
verbal description of the requirements for its operation – its “story.” At this
time, transforming the verbal story into the formal ELF is in the stages of
exploratory research.
We must declare the
domain of application for which we need the intelligent controller. A
distinction is made between closed and open systems. Closed systems can be
characterized by clear assignments of the problem to be solved and the ability
to construct a complete list of concrete user specifications in terms of
measurable variables. On the other hand, in open systems, the problem is not
totally clear, its parts are not concrete, the variables are not all listed at
the beginning of the design process, and the methods of observation and
registering are not outlined. Intelligent systems based on ELFs are appropriate
for the open system problem.
Defining the success of a system does depend
on the expectations of the customer. In Albus’ definition “success” is an
appropriate word, because it becomes a source of the emerging gradations in
intelligence.
Looking at the functioning of a system from
the outside, we can devise degrees of intelligence, which are linked to the
specifications of the controller. For example, we can define six degrees of
intelligence that separate an air conditioning unit from an artificial climate
control system.
1st degree of intelligence. An air conditioning
unit only knows a threshold for when to turn the compressor “on” or “off.” Even
the value of “accuracy” is frequently irrelevant for these systems.
2nd degree of intelligence. The system may be
required to stay within some interval of temperatures, with a given accuracy,
as well as some interval of humidity, with a given accuracy. The goal pursued
by the system is not a single goal-state, but is rather a zone determined by an
external function (in this case, based on human preferences).
3rd degree of intelligence. The system in the
previous example is now able to “learn” the human’s preferences, perhaps based
on the number of adjustments the human makes to the desired temperature and
humidity, and hence adjust the goal state. If the system learns the new goal
states quickly, it may reduce the need for human interference, and be
considered more successful.
4th degree of intelligence. An even more
interesting situation might happen if there is more than one user, and
different users have different policies of tuning the system. The Artificial
Climate System that would minimize the total number of cases of human
interference could be considered a system for achieving consensus in a
particular multi-player game.
5th degree of intelligence. A further development
of the system might be required if the owner of the unit in a hotel may want to
reduce the cost of energy required for keeping the customers happy. The system
can be designed so as to learn how to keep the average number of customer
complaints at a minimum, while minimizing the energy consumption. A high degree
of autonomy in this system allows it to assign the schedule for the functioning
of its subsystems.
6th
degree of intelligence. The previous systems
are subserviently autonomous, i.e., they control their own behavior, but the
goals are determined from the external user. If the system has a concept of self, it will try to keep all the
“customers” satisfied, while being concerned about its own lifespan, reducing
aging, increasing reliability, and so on.
The necessity for
gradations in the measurement of success of the intelligent system leads to
gradations in the measurement of the components of the ELF.
For the world modeling,
or knowledge representation, we can look at parameters that are meaningful,
such as the competence of the knowledge base (can it answer questions that they
rest of the system poses in order to do its job?), the memory depth, the number
of information units that can be handled, the number of levels in the system of
representation, the number of associative links between units of information,
and so on. For Behavior Generation, we can assess parameters such as the
horizon of planning at each level or resolution and the size of spatial scopes
of attention. For Sensory Processing, parameters to be considered include the
depth of details taken into account during the processes of recognition at a
single level of resolution, the number of levels of resolution that should be
taken into account during recognition and the minimum distinguishability unit
in the most accurate scale.
Further development of
the analysis of intelligent controllers requires construction of inner ELFs for
each subsystem of the main ELF (see Fig. 3).
7.
CONCLUSIONS
We have described how
the field of Intelligent Control has several underlying characteristics that
define its systems. We believe that the issue of control is just part of the
overall intelligence in a system; hence we expand our scope to look at
Intelligent Systems. In the broader scope, we can look at the capabilities of
the overall system, as gauged against the user’s expectations, in order to
extract definitions of intelligence. We can also consider how to model the
intelligent systems based on hierarchies of Elementary Loops of Functioning.
The parameters for the elements of the ELFs can be subject to measurement and
analysis. The measures of both the overall system’s capabilities and the parameters
of the constituent modules can be used to form the basis for a more formal
theory of intelligent systems and the study of metrics for such systems.
8.
REFERENCES
Ø J.
Albus, C. R. McLean, A. J. Barbera, M. L. Fitzgerald, "Hierarchical
Control for Robots and Teleoperators
Ø J.
Albus, "Outline for a Theory of Intelligence,"
Ø J.
S. Albus, A. M. Meystel, Behavior
Generation in Intelligent Systems
Ø P.
Antsaklis, “Defining Intelligent Control,”
Ø P.
Chiacchio, S. Chiaverini, and B. Siciliano, “User-Oriented Task description for
Cooperative Spatial Manipulators: One-Degree-of-Freedom Rolling Grasp,”
EmoticonEmoticon