 |
|
"Human
Supervision and Control in Engineering and Music" |
Dr.-Ing. Leon Urbas
Real time Dynamic Decision Making
in Supervisory Control
Abstract
The paper outlines some attributes of dynamic human-machine systems
which
are relevant to classifiy them as real time dynamic decision making
systems
from the persepctive of the human supervisory operator.
Characteristics of Dynamic Human Machine
Systems
Our research has its focus on modelling of cognitive behaviour of human
operators in dynamic human-machine-systems. The class of systems we are
looking at can be characterised by the following attributes:
- Dynamic system: the technical system shows a characteristic
dynamic behaviour,
which is not fully changeable due to some limitations of the system.
The
future development of the technical system without outer influence is
determined
by some memory in form of energy, mass or information storage.
- coupled multiple inputs multiple outputs: Single manipulated
variables
(input) of the system show influence on more than one observable state
variable of the system. This can be the cause for conflicting goals.
For
instance, raising the throughput in processes of the chemical
industries
often shows a contra productive effect in product quality.
- latent variables: The knowledge of manifest (direct observable)
variables
is not sufficient to interpret current or anticipate future behaviour
of
the system. Instead it is necessary to deduce latent variables from
observation
of manifest and manipulated variables.
- time variant dynamics: endogenous disturbances or deliberate
changes in
topology of the system may have great influence on the dynamic
behaviour
of the system due to new or fading interactions between different parts.
- open: the technical system is affected by exogenous disturbances
which
most often are not direct observable - it is necessary to deduce them
from
unusual behaviour of the system.
- real time: activities or sequels of activities have to be
executed until
certain deadlines to reach the intended goal. There are no means to
stop,
freeze or rewind the technical system.
The characteristics mentioned above, especially the ad hoc unknown
latent variables and the exogenous disturbances make the
decision-making
problem ill-defined: start and end of the problem are unknown and may
change
during the problem solving process. Due to the real time
characteristics
the time available is limited. Latent state variables and internal
coupling
of variables complicates the acquisition of accurate knowledge about
the
system. In consequence only limited time and uncertain knowledge is
available
to the responsible operator to judge about new situations, make a
decision,
and put goal oriented activities into execution. The mental models in
such
task environments, that can be deduced from learning through
interaction
and observation are generally only partial homomorphous, i.e. we
assume
that the relevant structure of the system can be mapped only partially
on the mental model. This makes sense from an economic perspective: the
requirements for a functional mental model which is useful for the
control
of a single variable are fundamentally different from the requirements
for a structural qualitative mental model which helps in failure
diagnosis.
Decision Making under Pressure of Time
Rational behaviour in real time dynamic decision-making systems in the
sense of good adaptation to the task environment makes it necessary to
revert to strategies, which reduce the need for time and cognitive
processing
resources. We assume that generally strategies with low execution time
and low demand for cognitive resources are chosen, as long as the
subjective
necessary power of anticipation can be reached. How effort and power of
anticipation may be
represented or calculated in a cognitive architecture is not clear
at the moment and object of research. To clarify things, some examples
for generic strategies are sketched which differ in their need of time
and the demand for cognitive resources:
- a priori strategy: The current state is anticipated by the mental
model
and recent observations, and may be followed by an activity where some
critical variables are compared.
- a posterior strategy: The current state is reconstructed from
observation
of current data over some period and may be some historical data. This
passive strategy may be coupled with well-directed manipulation of some
variables.
- erratic strategy: interact by random and hope for the best.
It depends on the task environment, whether a strategy is successful
and
in this sense adequate. The strategy of choice may be influenced by the
structure of the domain, the task itself, the necessary level of detail
and the engineering design of the supervisory control systems, i.e. the
task sharing between automation and human operator as well as the
design
of the interface.
Operator Modeling and Music
If we want to compare the conductor and his orchestra with the
supervisory
operator and his technical system from the perspective of common
cognitive
models, the author believes, that it is necessary to have a close look
on the
characteristics of the dynamic task environment, the tasks, the problem
solving strategies and the actions which can be compared or have to be
distinguished. The workshop is a highly welcome opportunity to start
with
this task. Some questions, which arise from current considerations,
are:
Is every orchestra able to play any score? If not, what are the limits?
Is conducting a real time decision making task? What are musicians
doing,
when the score assigns their instrument to pause?