![An Event Driven Parallel Blackboard System Modules emit events to the control unit or modify the blackboard data. Main sorts of events are modication and signal. A modiication event implies an access to the domain blackboard.With signals events, modules communicate their state to the control unit. Events only contain immediate information. This distinction between data ows and control ows is also made in non-blackboard agent architecture as 1]. The control unit receives events from modules and emits control signals to them. Common types of control signals are activation signal, inhibition signal. Modules which have all their conditions validated are activated by an activation ow. Inhibition signals trigger exceptions processing in the modules. The control unit is application independent. It is built and formalized as in 7].](https://www.researchgate.net/profile/Yves_Demazeau/publication/2670272/figure/fig2/AS:651183366623233@1532265621711/An-Event-Driven-Parallel-Blackboard-System-Modules-emit-events-to-the-control-unit-or_Q320.jpg)
Content uploaded by Yves Demazeau
Author content
All content in this area was uploaded by Yves Demazeau on Aug 04, 2014
Content may be subject to copyright.
Building Real Time Agents
using Parallel Blackboards and its use for Mobile Robotics
Michel Occello
, Yves Demazeau
y
LIFIA/IMAG/CNRS
46, avenue Felix Viallet
38031 Grenoble Cedex FRANCE
Michel.Occello@imag.fr - Yves.Demazeau@imag.fr
Abstract :
Vehicle intelligent control is a robotics real time
application where elaborated reasoning process and reactive process
work together and have to cooperate or more to be interdependent.
Multi-agent systems are well suited to such complex systems speci-
cation, because of their software engineering, their reasoning capa-
bilities (from a cognitive point of view) or their performances (from
a reactive point of view). This paper prop oses a model of agent in-
tegrating both reactive and deliberative capabilities adapted to real
time context. The paper introduces and discusses the use of a parallel
blackboard architecture to support the agent model in order to meet
real time constraints. An illustration of the functioning of the agent
architecture is given through a roadway trac scenario emphasizing
the main aspects of real time distributed decision making.
1 Introduction
Most real time systems can b e seen as realising three
steps : data acquisition, data processing and output
providing to human-machine interfaces or physical de-
vices. Real time intelligent systems for process control
or operator assistance aim to supervise external physi-
cal devices with the help of computers. They have to
work in relation with the real world. Agents systems
are well suited to applications in this domain, because
they oer modularity, exibility, and concurrency. But
in order to realise real time applications, agent systems
have to be improved to take into account the evolution
of the corresponding real environment.
Agents have to modify their behavior. They have to
be able to adapt their plans according to the dynamic
modication of the real world. To build real time multi-
agent systems, we have to integrate in a same agent,
cognitive capabilities (symbolic reasoning, social behav-
ior) to ensure the best cooperation between agents, and
Ma^tre de Conferences a l'Universite Pierre Mendes-France
(UPMF)-Grenoble II
y
Charge de Recherches au Centre National de la Recherche
Scientique (CNRS)
reactive capabilities to follow the evolution of the envi-
ronment as shown in [2].
We propose in this paper a mo del of an agent inte-
grating both reactive and cognitive asp ects. This agent
architecture is supported by a generic parallel black-
board model which supplies a powerful architecture to
implement agents. In the rst part, we give an insight
to some previous work that denes requirements for a
cognitive /reactive agent and proposes elements for an
agent model. Then we discuss this approach. In the sec-
ond part, the use of parallel blackboard architecture for
this type of agent is argued. The proposed architecture
of cognitive/reactive agent is detailed in the third part.
Finally, the activity of the resulting parallel blackboard
model of agent is exposed in the last section, in a real
time scenario of roadway trac.
We conclude by some perspectives ab out the use
of parallel and distributed blackboards to build agent
systems and about the integration of such blackboard
agents in a multi-agent integration platform.
2 Previous work
In [2], S. Bussman and Y. Demazeau have shown that
especially in complex domains neither purely reactive
nor purely cognitive approaches suce to meet the re-
quirements imposed by the environment.They extract
the main requirements to integrate reactive and cogni-
tive aspects in a same agent, and propose sp ecications
for a reactive / cognitive agent. We emphasize in the
rst part the requirements they established and then we
give an insight to the model they have proposed.
2.1 Requirements
Cognitive agents in a multi-agent world employ very
general, and therefore powerful metho ds, but suer of
slow algorithms, due to their complexity. Furthermore,
these algorithms, as they are designed, do not take into
account unpredicted events, apart from execution fail-
ures which are only detected but not evaluated. These
two severe drawbacks for cognitive agent systems are
exactly the advantages of the other reactive agents in
a multi-agent world : reactivity and fast action selec-
tion. On the contrary, it seems dicult to encode com-
plex behaviour into a rule system that is only based on
perception. Consequently, one would like to incorpo-
rate the advantages of both into a single model. In the
following, the requirements stated for simple reactive-
cognitive integration are listed.
Reactivity:
An agent should be able to
react
upon
unpredicted
events by adapting its b ehaviour. The re-
activity of an agent is located in the interpretation of
its environment. Unpredicted events have to b e recog-
nized before one can react. The recognition is followed
by an evaluation which determines the internal changes
that the agent has to take in order to react to the new
situation accordingly.
Adaptation of Behavioural Speed:
An agent should
be able to adapt its behavioural speed to the evolution
of the system. This requirement covers two aspects.
First, perception should b e suciently fast in order to
take a consistent picture of the system's state. Second,
the computation and the execution of the appropriate
action should be accomplished before ongoing changes
make it obsolete. Reactive models show very fast be-
haviour by the use of rule systems. As stated in our
requirement, we would like an agent not only to be fast
in his actions but also adaptive to the rate of change of
the system.
Symbolic Representation and Reasoning:
An
agent should be able to manipulate explicitly its knowl-
edge about the universe. This oers us the classi-
cal techniques of AI, such as deduction, planning, and
learning. These have been used by all cognitive models
as a basis for an agent. But their advantage of gener-
ality is compensated by the absence of reactivity and
of adaptation which we have shown to be indisp ensable
requirements for multi-agent systems.
2.2 Elements for a model
It will be the goal of the model proposed to accom-
plish the integration of the three requirements formu-
lated above into a consistent agent model.The task of
an agent model is to describe how the input data is pro-
cessed in order to determine the output of the agent
(see gure 1). In the model the scope of the input data
is restricted to digitized information which are called
sensory input
. Analogously, the output of an agent is
called
actions
. From this p oint of view, the terms
per-
ception
and
execution
refer to the processes necessary
to interact with the environment.
evaluator
goals
guards
triggers
actions
supervise
world model
create, suspend
kill
replan set
perception
sensory input
Decision
Capabilities Capabilities
Reasoning
schedule
Capabilities
Commitment
Figure 1: Agent Model
This previous model is the base of the agent model
presented in our paper. Its main aspects are discussed in
the next part where we propose improvements to obtain
an adapted agent architecture.
2.3 Discussion
Adding communication.
This work aims to describe
how the input data is processed in order to determine
the output of the agent. It restricts the scope of the
input data to information called sensory input obtained
from the environment. We have to take into account
in the same way the information from the other agents.
Analogously, actions are supposed to interact with the
environment, the model has to consider the interactions
with the agents too.
Real time agent control.
Our objective is to extend
this model to obtain a mo del really applicable to real
time systems. A real time agent can be viewed as a
process control system that aims to supervise a physi-
cal device in its environment. The state of the physical
device is expressed through a set of measures and events
obtained from sensors. This state is continuously mod-
ied, witness to the system evolution. The sup ervision
of process consists in recording and maintaining a view
of the controlled system (for an operator for example)
according to the modications of the state. If the con-
trol system has to pilot the physical device, it has to
react to these modication by emitting orders to physi-
cal eectors. Resp onse time must be viable with regard
to the evolution of the controlled device.
Stressing the reactive behavior.
As opposed to
transformational
program which have to provide results
from input data and evolve at their own rythm, real
time programs have to b e
reactive
i.e. sub ject to the
evolution of their environment to which they are related
and which x their cadence. Unfortunately, control sys-
tems (and specially knowledge based ones) cannot b e
totally situated in these classes, because they use trans-
formational programs (algorithms) called by a reactive
kernel.
Validation through the organisation.
The previ-
ous mo del proposes a collection of transformational and
reactive modules, but doesn't explicit which organisa-
tion between these mo dules can guarantee a satisfying
global real time working. At this sense, a centralized
control approach seems to be better adapted, it is the
paradigm proposed by such systems. It needs to intro-
duce a synchronous control unit which ensure reactivity
in the decision making caused by an evolution of the
environment [7].
Parallelisation and synchronisation.
However, us-
ing agents which interact in real time with an environ-
ment pose problems of parallelism and synchronization
in their structures. Parallelism is an imp erative need
for a real time system :
- either because the physical device consists of dis-
tributed equipments,
- either for security, reliability or eciency.
In summary, we will propose in the following sections
to work towards a complete model of agent, integrating :
- A symbolic representation of b oth the environment
and the other agents,
- A sensing mechanism for both environment and
agents events,
- A graduated adaptation of the behaviour in reaction
to these events,
- An organisation for this agent supplying parallelism
and reactivity mechanisms in order to satisfy real time
constraints.
A parallel blackboard model is prop osed to achieve
this purpose. This choice is argued in the next section.
3 Why a parallel blackboard to
support a real-time agent ?
This section presents a basic parallel blackboard model
initially prop osed in [7]. We will later discuss its adap-
tation to requirements for a real time agent.
A parallel blackboard model :
Classically, a blackboard system consists of a data
structure called blackboard, several modules, a control
mechanism.
A parallel blackboard architecture aims to really ex-
press the inherent parallelism of the conceptual black-
board model [3].
Modules react towards modications of the black-
board, for their activations and inhibitions. They work
on a local context which is a part of the blackboard data.
The blackboard contains domain data (used for problem
resolution) and control data (summary of the state of
the resolution). A control mechanism is in charge of
the communications between modules and of control of
the management of modules activity
(gure 2).
Modules
Control
Unit
Control
Data
Domain
Data
Blackboard
Event Stream
Data Exchange
Control Stream
Event
Figure 2: An Event Driven Parallel Blackboard System
Modules emit events to the control unit or modify
the blackboard data. Main sorts of events are
modi-
cation
and
signal
. A modication event implies an
access to the domain blackboard.With signals events,
modules communicate their state to the control unit
.
Events only contain immediate information. This dis-
tinction between data ows and control ows is also
made in non-blackboard agent architecture as [1]. The
control unit receives events from modules and emits con-
trol signals to them. Common types of control signals
are
activation signal
,
inhibition signal
. Modules
which have all their conditions validated are activated
by an activation ow. Inhibition signals trigger excep-
tions processing in the modules. The control unit is
application indep endent. It is built and formalized as
in [7].
Adequation with requirements :
A complete study of the use of blackboard systems for
dynamic systems control [7] proposes a few privileged
points to adopt the blackboard concept to organize a
real time agent :
1.
A real time system uses for many reasons a dis-
tributed architecture : it seems necessary to adopt
a
parallel blackboard model
. A parallel blackboard system
uses a functional decomposition of the control of sys-
tems where modules represents dierent tasks of organs
of the physical device, as control functions of eectors
or access to sensors constituting a heterogeneous archi-
tecture.
2.
A blackboard system is no longer seen as a trans-
formational system but as a reactive system :
decision
making aims no longer to nd the best solution in a lim-
ited time but to respond to an event in a limited time
which depends of the context
.
3.
The evolution of the state can question the activity
of some modules :
modules must be sensitive to interrup-
tion signals
. The blackboard becomes a dynamic tool
which coordinates and schedules mo dules and ensures
synchronization.
4.
Parallelism allows by the asynchronous execution
of modules
to acquire data in real time through special-
ized modules that constitute the interface with the real
world
. The software manages the activity of physical
modules (directly related to process) and logical ones
(for the system evolution) through the data structure
representing the world. The structure represents the
state of the environment and of the controlled system.
Modules modify the state of the structure which evolves
and oers at each time the supervision of the process.
5.
The control mechanism must choose modules
whose activity can be performed in the right delay,
but
time-out tests must be provided to suppress mod-
ules whose duration are too expensive according to their
deadline
. Separation between the blackboard control
mechanism and the domain activity oers opp ortunities
for action scheduling under temp oral constraints.
6.
Control complexity must be reduced by the
elimi-
nation of competing interpretations and reliable data l-
tering
. Recording the modication of the environment
on the blackboard and the choice of action in function
of the state of the blackboard provides an interesting
mechanism for reactivity.
This management, control mechanism, must be trans-
parent as in any programming systems. Such an inter-
pretation of the concept is used for sensors fusion and
multi-robots coordination.
4 A model of agent integrating
cognitive and reactive aspects
The integration of cognitive and reactive capabilities is
possible only with the use of parallelism in the structure
of the agents. Separation b etween Decision/Reasoning
and Perception/Communication tasks allows a contin-
uous supervision of the evolution of the environment.
The reasoning model of our agent is based on the
Goals Action
Plans
Planning Scheduling Executing
Decision
Universe
Figure 3: Control Process of the Agent Mo del
Perception/Decision/Reasoning/Action paradigm. The
cognitive reasoning is so preserved. Predicted events
contribute to the normal progress of the reasoning pro-
cess. Decision modules evaluate the imp ortance of the
unpredicted events and the obligation to place new goals
on the mental model. New goals imply the activation of
reasoning modules to partially or totally replan accord-
ing to the importance of the event. The agent control
process can be explicited by the gure 3.
We describe now the dierent modules needed by a
cognitive/reactive agent proposed by [2], organized ac-
cording to the blackboard model presented b elow. The
System
Protocol
Operating
Environment Others Myself
State of the
Others Actions
Plan
Goals
Protocol
Perception 1
PERCEPTION
Interpretation
Communication
COMMUNICATION
UNIT
Control Flows Data Exchange
Action n
Action 1
ACTION
Evaluator
AGENT CONTROL REASONING
Planner Scheduler Executor
DECISION EXECUTION
CONTROL
Figure 4: A cognitive/reactive parallel blackboard agent
centre of the agent is its
world model
. This model
comprises its knowledge ab out
the environment, the
mental states of other agents, and its own mental
state
. The proper mental state includes in particular
the plans that are being executed or which the agent
takes into consideration. This mo del is maintained by
an interpretation process of the sensory data. This
model constitutes the
domain blackboard
. Evolving
in a real world, each agent must integrate capabilities
of perception realized by sensor devices. The knowledge
about the environment is build by
PERCEPTION
modules.
To ensure the reactivity of the agent, an evaluator
continuously examines this world model.
AGENT
CONTROL modules
detect situations to which the
agent needs to react, evaluates them, and decides to take
the appropriate actions which may be of the form to cre-
ate, suspend, or kill goals, i.e. to change the context of
the planning and executing process. The continuous su-
pervision of the agent's situation ensures that the agent
can react to unpredicted events at any time. The same
mechanism takes into account interaction with other
agents using the interaction protocols proposed by [4]
(
COMMUNICATION mo dules
). Furthermore, the
planner has the p ossibility to include internal actions,
such as replanning or the setting of
guards
and
triggers
,
in order to be more adaptive at run time. Guards and
triggers, supply information ab out the current situation
during the execution phase of actions, they are stored
in the domain blackboard.
Whenever a goal is created (or modied) a plan is
searched that achieves the goal, this task is realized
by
REASONING modules
. Plans relative to each
known goal are stocked in the part of the blackboard
concerning the mental state of the agent. The planner
details the action in the order in which they will be
performed. This process may be guided by hierarchical
planning that attempts to infer the action sequence in
a top-down fashion. In simple applications, we can as-
sume that the agent have plans for all encountered goals
and p ossesses all needed actions; the planning process
is then reduced to a fast pattern-matching algorithm.
Constructed plans are scheduled consistently in the
blackboard. Due to this technics, if the agent has to act
fast, the scheduling and the execution of an incomplete
plan can start before the planning pro cess is nished.
This ensures the adaptation of the agent to the evolu-
tion speed of his environment and is necessary if the
agent pursues several goals at the same time. The com-
mitted actions are performed at the scheduled time by
the
ACTION mo dules
.
All modules are managed by the control unit, the
global model is presented in gure 4. This multi-
modules approach allows a modular and independent
description of each of the action and perception tasks
in separate modules.
5 Application to mobile robotics
We illustrate in this section how runs the agent architec-
ture by an application to mobile rob otics agents. The
next situation example emphasizes the agent organiza-
tion and its behavior in front of trac. They show the
integration of deliberative and reactive aspect of the
agent decision making according to the progressive de-
cision process exposed in the previous section.
Let us assume a robotic agent on a road. In order
to plan or to achieve symbolic decision making tasks,
This agent needs cognitive capabilities. Thus it must
have a representation of the environment. But as it
must control a physical device (the vehicle) in a real
time scenario, this representation must be dynamically
maintained, to resp ond to unpredicted events.
A normal working mode consists in the planning of
trajectory in function of a goal. A goal is expressed by
a nal position, sp eed and orientation for the robotics
mobile device on the road map. A nominal path is cal-
culated by a dedicated module and placed in a goal tree.
According to the model of agent proposed in the pre-
vious section, we can describe an agent structure for this
application. The
Planner
Module is triggered by a new
goal (i.e by the reception by the blackboard control unit
of an
Event(New Goal)
signifying a modication on the
shared blackboard). Part of Reasoning modules, it aims
to nd a plan to satisfy a goal. Plans (list of actions)
are stored on the blackboard.
The
Scheduler
Module chooses a plan, takes all ac-
tions of this plan and schedules them at the required
place in the
BB:agenda of actions
. If the plan cannot
be scheduled, a new Event(New Goal) can propose to
replan.
In the case of a normal moving, nominal path is pro-
duced by a
ACTION:Plan Trajectory
Module. This
module computes a nominal trajectory for a nominal
speed-time prole from the current posture, the nal
goal and the static model of the environment stored in
the
BB:environment
part of the blackboard. The
AC-
TION:Execute Trajectory
Module generates commands
to the physical device each sample period of time t.
According to the potential eld metho d [6], a command
consists in the nominal path parameters modied by
potential eld parameters. These two actions modules
are executed by the blackboard control unit by reac-
tion to Event(Activate Action ...) from the
Executor
Module. Once the execution phase is started, the agent
is in a perception/action mode, since a plan is b eing
performed. The perception modules react to dynamic
modication of the environment.
In the situation represented by gure 5, the agent
follows a vehicle with a lower-speed. The vehi-
cle is detected by the perception modules whose re-
sults trigger a
AGENT CONTROL:Evaluator
Mod-
ule. The trigger Same way vehicle present is vali-
dated, and implies to Acquire vehicle parameters. This
tasks is done by the
ACTION:Agent Interaction
Mod-
ule to ask the other agent its speed, direction, etc.
The
COMMUNICATION
Modules manage the mail-
VV1
AA1
Scope of Agent A Sensing
Overtaking Trajectory
Figure 5: Situation of Perception-Communication/ De-
cision/ Action : Lower-Speed Vehicle on the Road
box and emit Event(New Messages). A
AGENT
CONTROL:Evaluator
evaluates consequences of the
message reception (eventually according to a pro-
tocol). It mo dify the
BB:Others:vehicle parameters
with exchanged values, update the
BB:New Adaptations
elds and emit Event(New Adaptations) resulting
of the evaluation of
BB:communication triggers
and
BB:communication guards
. The
DECISION:Scheduler
Module is triggered by a necessary adaptation, it sched-
ules the action stored in the blackboard with an im-
mediate date of activation, so that the executor emit
immediately the
Event(Activate Action...)
. The action
can b e
ACTION:Slow Down
or
ACTION:Overtaking
which will modify the potential parameters of the com-
mands stored in the
BB:Goal Tree
. This behavior fol-
lows the Perception/Decision/Action process without
replanning.
This situation illustrates, so, the functioning of the
parallel blackboard architecture of agent presented in
this paper. Both perception of static environment and
interaction with other agents are taken into account. All
stages of the decision process are proposed, allowing a
really ecient achievement of the integration of reactive
and deliberative capabilities in a real time context.
6 Conclusion and Perspectives
Viewing Parallel and Distributed Blackboards
as Support Architectures for Multi-Agents Sys-
tems
Blackboard Approach brings software exibility
and modularity the implementation of Action and Per-
ception/Communication tasks. Furthermore, the inde-
pendence between modules allows the coupling of both
simulated and physical control modes of activity.
The study of a parallel blackboard model specially
adapted to reactivity brings a powerful support for
problems involving both centralized representation of
data (as cognitive agents) and reaction to unpredicted
events (reactive aspects).
A multi-agent system will be in fact implanted as a dis-
tributed hybrid blackboard system, using a set of par-
allel blackboard systems in a distribution close to the
DVMT structure [5].
Integration in the MAGMA platform
A generic
tool has been developed in C++ using UNIX com-
munication libraries [8]. It oers a graphical inter-
face to build parallel blackboard systems and so reac-
tive/cognitive agents. This software is now under devel-
opment using DPSK+P libraries of active objects. The
model of agent will be so integrated with its tool into
the MAGMA platform of development and simulation
of multi-agent systems, currently studied in our group.
7 Acknowledgements
The author would like to thank Stephan Bussmann for
his study on the integration of cognitive and reactive
aspects in a agent structure, which is partly at the origin
of this agent model.
References
[1] O. Boissier and Demazeau Y. A multi-agent con-
trol architecture for studying the control of an inte-
grated vision system. In
IEEE Int Conf. on Multi-
Sensor Fusion and Integration for Intelligent Sys-
tems - MFI'94
, Las Vegas, USA, october 1994.
[2] S. Bussmann and Y. Demazeau. An agent model
combining reactive and cognitive capabilities. In
To
appear in Proccedings of IEEE International Confer-
ence on Intel ligent Robots and Systems - IROS'94
,
Munchen, September 1994.
[3] D.D. Corkill. Advanced Architectures: Concurrency
and Parallelism. In V. Jagannathan, R. Dodhiawala,
and L.S. Baum, editors,
Blackboard Architectures
and Applications
, chapter II, pages 77{83. Academic
Press, 1989.
[4] Y. Demazeau, O. Boissier, and J.L. Koning. Inter-
action protocols in distributed articial intelligence
and their use to control robot vision systems. In
IEEE Int. Conf. on Systems, Man and Cybernetics
,
San Antonio,Texas, USA, october 1994.
[5] E.H. Durfee.
Coordination of Distributed Problem
Solvers
. Kluwer Int. Series in Engineering and Com-
puter Science, 1988.
[6] T. Fraichard and Laugier C. On line reactive plan-
ning for a non holonomic mobile in a dynamic
world. In
Proc. of IEEE International Conference
on Robotics and Automation
, pages 432{437, Sacra-
mento, Ca., april 1991.
[7] M. Occello. Distributed and Parallel Blackboards :
Application to Dynamic Systems Control in rob otics
and Computer Musics.
PhD Thesis Report, Uni-
versity of Nice - Sophia Antipolis
, january 1993 (In
French).
[8] M. Occello and M. C. Thomas. A Parallel Black-
board Generic Tool for Intelligent Robotics. In
Proc.
of IEEE Tools for Articial Intel ligence
, Arlington,
VA, USA, november 1992. IEEE Press.
