Decision-Making in an Embedded Reasoning System

Michael P. George
y Francois Felix Ingrandz

Australian AI Institute Articial Intelligence Center
1 Grattan Street SRI International
Carlton, Victoria 3053 333 Ravenswood Avenue
Australia Menlo Park, California 94025
Abstract planning systems cannot operate under the stringent
constraints on both information availability and de-
The development of reasoning systems that cision time that are typical of real-time applications.
can reason and plan in a continuously As a rule, these systems formulate an entire course
changing environment is emerging as an of action before commencing execution of the plan
important area of research in Arti cial In- Wilkins, 1988]. However, in real-world domains, in-
telligence. This paper describes some of the formation about how best to achieve some given goal
features of a Procedural Reasoning System (and thus how best to complete a plan) can often
(PRS) that enables it to operate eectively only be acquired after executing some initial part of
in such environments. The basic system that plan.
design is rst described and it is shown Real-time constraints pose yet further problems
how this architecture supports both goal- for traditionally structured planning and reasoning
directed reasoning and the ability to react systems. First, the planning and deductive tech-
rapidly to unanticipated changes in the en- niques typically used by these systems are very time
vironment. The decision-making capabili- consuming. While this may be acceptable in some
ties of the system are then discussed and situations, it is not suited to domains where replan-
it is indicated how the system integrates ning is frequently necessary and where system viabil-
these components in a manner that takes ity depends on readiness to act. Second, traditional
account of the bounds on both resources planning systems usually provide no mechanisms for
and knowledge that typify most real-time responding to new situations or goals during plan
operations. The system has been applied to execution, let alone during plan formation. Yet the
handling malfunctions on the space shut- very survival of an autonomous system may depend
tle, threat assessment, and the control of on its ability to react to new situations and to mod-
an autonomous robot. ify its goals and intentions accordingly George and
Lansky, 1987].
1 Introduction A number of systems developed for the control of
robots and other real-time processes do have a high
While there has been an increasing amount of inter- degree of reactivity Brooks, 1985, Kaelbling, 1987].
est in the development of reasoning systems suited to Such architectures could lead to more viable and ro-
real-time operations, very few currently exhibit the bust systems than traditionally structured planning
kind of behavioral properties that we would expect systems. Yet most of this work has not addressed the
of such systems Laey et al., 1988]. Much of this de- issues of general problem-solving and commonsense
rives from their inability to reason eectively about reasoning the work is instead almost exclusively de-
actions and plans. On the other hand, most existing voted to problems of navigation and the execution
of low-level actions.
This paper appeared in the Proceedings of the There has been some recent work in the design
Eleventh International Joint Conference on Arti cial In- of systems that attempt to integrate goal-directed
telligence, Detroit (Michigan), August 1989. reasoning and reactive behavior. For example, the
y Also a member of the Arti cial Intelligence Center,
SRI International, and the Center for the Study of Lan- RAP system Firby, 1987] invokes tasks on the basis
guage and Information, Stanford University. of system beliefs about the current world situation
z Partially supported by the French research agency: and then expands these tasks hierarchically. It has
Institut National de Recherche en Informatique et en many features in common with the system described
Automatique (INRIA). in this paper. However, it does not appear to pro-
0 This research is partly supported by the National vide suciently powerful mechanisms for balancing
Aeronautics and Space Administration, Ames Research the decision-making requirements against the con-
Center, under Contract No. NAS2-12521. straints on time and information that are typical of
complex domains.
As Bratman, Israel, and Pollack 1988] remark: MONITOR
\... Rational agents must perform both means-end
reasoning and weigh alternative courses of action so ? 6
an adequate architecture of intelligent agents must DATABASE
(BELIEFS)
KA LIBRARY
(PLANS)
therefore include capabilities for both. The design
of such an architecture must also specify how these QQ  SENSORS 
capacities interact. But there is yet another prob- INTERPRETER
lem. All this must be done in a way that recognizes (REASONER) ENVIRONMENT
the fact that agents are bounded both in resources
QQ 6
and knowledge]." In this paper, we describe a sys-
tem architecture that aims to achieve this balance  EFFECTORS

between acting and decision-making. GOALS INTENTION

STRUCTURE 6
2 Procedural Reasoning Systems - GENERATOR
COMMAND
The problem domain we discuss in this paper is the
task of malfunction handling for the Reaction Con- Figure 1: System Structure
trol System (RCS) of NASA's space shuttle. The
shuttle contains three such systems, one forward and
two aft. Each is a relatively complex propulsion KAs. These will typically be current observations
system that is used to control the attitude of the about the world or conclusions derived by the sys-
shuttle. The astronaut handles system malfunctions tem from these observations, and these may change
by carrying out prespeci ed malfunction handling over time. The knowledge contained in the database
procedures. These procedures can be viewed as un- is represented in rst-order predicate calculus.
elaborated plans of action, and are designed to be State descriptions that describe internal system
executed in a complex and changing environment. states are called metalevel expressions. These typ-
To tackle this problem, we developed an embed- ically describe the beliefs, goals, and intentions of
ded reasoning system called PRS (Procedural Rea- the system, as well as other important processing
soning System). Early versions of the system have information.
been described before George and Lansky, 1986a, 2.2 Goals
George and Lansky, 1986b, George and Lansky,
1987] and the full application is described in a re- Goals are expressed as conditions over some interval
cent report George and Ingrand, 1988]. of time (i.e., over some sequence of world states) and
PRS consists of (1) a database containing current are described by applying various temporal opera-
beliefs or facts about the world (2) a set of current tors to state descriptions. This allows us to represent
goals to be realized (3) a set of plans (called Knowl- a wide variety of goals, including goals of achieve-
edge Areas) describing how certain sequences of ac- ment, goals of maintenance, and goals to test for
tions and tests may be performed to achieve given given conditions. A given action (or sequence of ac-
goals or to react to particular situations and (4) tions) is said to succeed in achieving a given goal if
an intention structure containing those plans that its execution results in a behavior that satis es the
have been chosen for eventual] execution (Figure goal description.
1). An interpreter (or inference mechanism) manip- As with state descriptions, goal descriptions are
ulates these components, selecting appropriate plans not restricted to specifying desired behaviors of the
based on the system's beliefs and goals, placing those external environment but can also characterize the
selected on the intention structure, and executing internal behavior of the system. Such descriptions
them. are called metalevel goal speci cations.
The system interacts with its environment, includ-
ing other systems, through its database (which ac- 2.3 Knowledge Areas
quires new beliefs in response to changes in the en- Knowledge about how to accomplish given goals or
vironment) and through the actions that it performs react to certain situations is represented in PRS by
as it carries out its intentions. declarative procedure speci cations called Knowl-
edge Areas (KAs). Each KA consists of a body, which
2.1 The System Database describes the steps of the procedure, and an invoca-
The contents of the PRS database may be viewed as tion condition, which speci es under what situations
representing the current beliefs of the system. Typi- the KA is useful (applicable). Together, the invoca-
cally, these will include facts about static properties tion condition and body of a KA express a declara-
of the application domain, such as the structure of tive fact about the results and utility of performing
some subsystems or the physical laws that must be certain sequences of actions under certain conditions
obeyed by certain mechanical components. Other George and Lansky, 1986a].
beliefs are acquired by PRS itself as it executes its The body of a KA can be viewed as a plan or
plan schema. It is represented as a graph with one tiple least elements (called the roots of the struc-
distinguished start node and possibly multiple end ture). An intention earlier in the ordering must
nodes. The arcs in the graph are labeled with the be either realized or dropped (and thus disappear
subgoals to be achieved in carrying out the plan. from the intention structure) before intentions ap-
Successful execution of a KA consists of achieving pearing later in the ordering can be executed. This
each of the subgoals labeling a path from the start precedence relationship between intentions enables
node to an end node. This formalism allows richer the system to establish priorities and other relation-
control constructs (including conditional selection, ships between intentions.
iteration, and recursion) than most plan representa-
tions. 2.5 The System Interpreter
The invocation condition contains a triggering The PRS interpreter runs the entire system. From
part describing the events that must occur for the a conceptual standpoint, it operates in a relatively
KA to be executed. Usually, these consist of the ac- simple way. At any particular time, certain goals are
quisition of some new goal (in which case, the KA is active in the system and certain beliefs are held in
invoked in a goal-directed fashion) or some change in the system database. Given these extant goals and
system beliefs (resulting in data-directed or reactive beliefs, a subset of KAs in the system will be appli-
invocation), and may involve both. cable (i.e., will be invoked). One or more of these
Some KAs have no bodies. These are so-called applicable KAs will then be chosen for execution and
primitive KAs and have associated with them some thus will be placed on the intention structure.
primitive action that is directly performable by the In determining KA applicability, the interpreter
system. will not automatically perform any deduction. Both
The set of KAs in a PRS application system not beliefs and goals are matched directly with invoca-
only consists of procedural knowledge about a spe- tion conditions by using uni cation only. This al-
ci c domain, but also includes metalevel KAs | that lows appropriate KAs to be selected very quickly
is, information about the manipulationof the beliefs, and guarantees a certain degree of reactivity. If we
desires, and intentions of PRS itself. For example, allowed arbitrary deductions to be made, we could
typical metalevel KAs encode various methods for no longer furnish such a guarantee. However, PRS
choosing among multiple applicable KAs, modify- is always able to perform any deductions it chooses
ing and manipulating intentions, and computing the by invoking appropriate metalevel KAs. These met-
amount of reasoning that can be undertaken, given alevel KAs are themselves interruptible, so that the
the real-time constraints of the problem domain. reactivity of the system is retained.
2.4 The Intention Structure Once selected, the chosen KAs are inserted into
the intention structure. If a selected KA arose due
The intention structure contains all those tasks that to the acquisition of a new intrinsic goal1 or a new
the system has chosen for execution, either imme- belief, it will be inserted into the intention structure
diately or at some later time. These adopted tasks as a new intention. For example, this will be the
are called intentions. A single intention consists of case for a KA that is invoked by the activation of
some initial KA together with all the sub-] KAs that some caution-warning alarm during shuttle opera-
are being used in attempting to successfully execute tions. Otherwise, the KA instance must have arisen
that KA. It is directly analogous to a process in a as a result of some operational goal of some existing
conventional programming system. intention, and will be pushed onto the stack of KAs
At any given moment, the intention structure may comprising that intention.
contain a number of such intentions, some of which Finally, an intention at one of the roots of the
may be suspended or deferred, some of which may intention structure is selected for further execution.
be waiting for certain conditions to hold prior to ac- The next step of that intention will comprise either
tivation, and some of which may be metalevel inten- a primitive action or one or more unelaborated sub-
tions for deciding among various alternative courses goals. If the former, the action is directly initiated
of action. if the latter, these subgoals are posted as new oper-
For example, in handling a malfunction in the ational goals of the system.
RCS, the system might have, at some instant, three Execution of primitive actions can eect not only
tasks (intentions) in the intention structure: one sus- the external world but also the internal state of the
pended while waiting for, say, the fuel-tank pressure system. For example, a primitive action may oper-
to decrease below some designated threshold an- ate directly on the beliefs, goals, or intentions of the
other suspended after having just posted some goal system. Alternatively, the action may indirectly af-
that is to be accomplished (such as interconnect- fect the system's state as a result of the knowledge
ing one shuttle subsystem with another) and the gained by its interaction with the external world.
third, a metalevel procedure, being executed to de-
cide which way to accomplish that goal. 1 An intrinsic goal is one that is not a means to an
The set of intentions comprising the intention already intended end, whereas an operational goal is a
structure form a partial ordering with possibly mul- subgoal of some already existing intention.
 At this point the interpreter cycle begins again: situation, the failure to achieve a certain goal, or the
the newly established goals and beliefs trigger new awakening of some previously suspended intention.
KAs, one or more of these are selected and placed In some of these cases (such as the latter two
on the intention structure, and nally an intention is of those mentioned above), these conditions will be
selected from that structure and partially executed. known at the beginning of each interpretation cycle.
Unless some new belief or goal activates some new But others (such as the number of KAs applicable
KA, PRS will try to ful ll any intentions it has pre- at a given moment) can only be determined part-
viously decided upon. This results in focussed, goal- way through this cycle. Thus, the actual manner of
directed reasoning in which KAs are expanded in a invocation of metalevel KAs is somewhat complex
manner analogous to the execution of subroutines George and Ingrand, 1988], but eventually results
in procedural programming systems. But if some in the selection of a single KA at some level in the
important new fact or goal does become known, metahierarchy. This KA is then executed just as
PRS will reassess its current intentions, and perhaps any other KA. In the process of execution, such met-
choose to work on something else. Thus, not all op- alevel KAs typically make choices about the adop-
tions that are considered by PRS arise as a result of tion of lower-level KAs (which may themselves be
means-end reasoning. Changes in the environment metaKAs), the posting of alternative subgoals, or
may lead to changes in the system's goals or be- the best way to respond to goal failures. Any of
liefs, which in turn may result in the consideration these decisions may require an arbitrary amount of
of new plans that are not means to any already in- processing or deduction, yet system reactivity is al-
tended end. PRS is therefore able to change its focus ways guaranteed (see Section 5.1).
completely and pursue new goals when the situation Thus, when decisions have to be made in the ab-
warrants it. In many space operations, this happens sence of any information about what is best to do,
quite frequently as emergencies of various degrees of the system interpreter defaults to some xed deci-
severity occur in the process of handling other, less sion procedure. However, if there is any knowledge
critical tasks. that can be brought to bear on the decision (via
We have, of course, left out a crucial component of appropriate metalevel KAs), this will override the
the system's reasoning: how does it make the various default action and determine the selection. In this
selections and decisions mentioned above? We now way, the addition of appropriate metalevel KAs en-
turn to this and other important features of PRS. ables the system to make more informed choices at
the cost of longer decision times.
3 Integrating Decision-Making and This approach to the design of PRS is important
Means-Ends Reasoning in another way. We can now construct a relatively
simple basic interpreter, knowing that the system
PRS is designed so that, in the absence of any de- is not going to always be constrained to behave in
cision knowledge being provided to the system, it the way dictated by this component | its decisions
nevertheless continues to function in an acceptable can, in situations determined by the invocation con-
way as an embedded reasoning system. As more and ditions of the metalevel KAs, be overridden. Of
more decision knowledge is added to the system, it course, it is important that the basic interpreter be
can more eectively choose its actions, but always able to handle appropriately the most commonly oc-
with a guaranteed upper bound on reaction time. curring situations so that the metalevel KAs are uti-
This is accomplished by embedding in the system lized only in the more exceptional circumstances.
interpreter certain xed decision-making processes Before concluding this section, it is important
that are stringently bounded in execution time, yet to note that the decision-making behavior of PRS
which can be overridden whenever the system can is strongly inuenced by the choice of the invoca-
bring more powerful decision-making knowledge to tion conditions of metalevel KAs. For example, if
bear. Bratman et al. 1988] have outlined a simi- these conditions are such that the decision-making
lar design philosophy, in which they distinguish be- metaKAs are frequently invoked, PRS will spend
tween a computationally ecient compatibility lter more time making decisions than otherwise. It will
and a possibly computationally complex lter over- thus tend to act in a cautious manner Bratman et
ride mechanism. In PRS, the xed decision-making al., 1988], carefully choosing what actions to per-
processes are hardwired into the basic system inter- form next. If, on the other hand, these metalevel
preter, whereas the knowledge to override or elabo- KAs are rarely invoked, PRS will act in a bold man-
rate these decisions is contained in appropriate met- ner, rapidly choosing its actions in response to the
alevel KAs. changing world in which it is embedded.
The way these metalevel KAs are brought to bear
on any particular problem is via their invocation cri-
teria. These criteria may depend both on conditions 4 The Nature and Role of Intentions
obtaining in the external world and, more typically, Intentions play a signi cant role in PRS. In this sec-
on conditions relating to the internal state of the tion we examine their use both in managing events
system. Such conditions might include, for exam- in the real world and in limiting the amount of de-
ple, the applicability of multiple KAs in the current cision making that has to be undertaken, given the
real-time constraints of the domain. possible to continually reassess one's plans of action.
What makes the approach workable is that the basis
4.1 Intention States upon which one chooses a particular plan of action
In PRS, an intention can be in one of three possi- is more often correct than not.
ble states: active, suspended, or conditionally sus- It is not only in means-ends reasoning that PRS's
pended. An active intention is one that is available commitment to its existing intentions is important.
for execution as soon as it becomes a root of the For example, in tackling some new task, it is of-
intention structure. A suspended intention is one ten desirable that the means or time chosen for ac-
that PRS has adopted but for which no decision has complishing that task take account of one's exist-
been made as to when it should be carried out that ing intentions towards the ful llment of other tasks.
is, it must be explicitly activated before it can be In the RCS application, this happens, for example,
executed. A conditionally suspended intention (or, when PRS receives a request for a pressure reading
simply, a conditional intention) is one that is tem- when it is in the process of evaluating the status
porarily suspended until some speci ed condition, of a suspected faulty transducer. In this case, PRS
called the activation condition of the intention, is will either defer or suspend attending to that re-
satis ed. quest (possibly advising the requester) until it has
In PRS, intentions can be suspended (condition- completed its evaluation of the transducer.
ally or otherwise) by means of certain metalevel
KAs. When a conditional intention is reactivated, 4.3 The Establishment and Dropping of
it is necessary to decide whether or not to reorder Goals
the intention structure and what to do, if anything, One does not want to be committed to one's inten-
with currently executing intentions. In the absence tions forever. Thus, it is important to understand
of any decision knowledge, it is desirable to ensure the way in which intentions are managed in PRS |
immediate response to the condition that activated in particular, how the goals comprising an intention
the intention. Thus, the basic interpreter is designed are established and dropped.
to begin execution of the reactivated intention im- As discussed in previous sections, intentions are
mediately, suspending all currently executing inten- established by a change in the system's goals or
tions. However, as with all other default decisions beliefs, and their execution will generate certain
taken by the interpreter, metaKAs can respond to operational goals. Should the system attempt to
reactivation and choose to override this decision. achieve an operational goal and fail, that goal will be
Similarly, various metalevel KAs exist for reorder- reestablished, and another attempt made to achieve
ing intentions on the intention structure. In the ab- it. This will continue until the system comes to be-
sence of any decision knowledge the basic interpreter lieve either that the goal is accomplished ( through
inserts new intentions at the root of the existing its own eorts or those of some other agent) or that
intention structure, and removes them when com- the goal cannot be readily accomplished. Once this
pleted (see Section 4.3). However, metalevel KAs state is reached, the goal will be dropped.
may choose to alter this ordering on the basis of This raises two related issues: what other at-
task priorities, time availability, and so on. tempts are made to achieve the goal and how does
the system come to believe a goal cannot be readily
4.2 The Role of Intentions accomplished? We shall answer the latter question
The role of commitment to previously adopted rst. One way for the system to believe that a goal
plans or intentions has been claimed to be a criti- cannot be accomplished (or, at least, that the goal
cal component of rational agency Bratman, 1987, is not worth pursuing further) is to deduce it using
Cohen and Levesque, 1987]. It is also a critical com- appropriate metalevel KAs. However, the system
ponent of the PRS architecture, allowing the system rarely has sucient knowledge of the world to prove
to meet the real-time constraints of a continuously that a goal cannot be achieved this is thus unaccept-
changing world. able as the only (or even as the primary) manner of
Unless some particular metalevel KA intervenes, dropping goals (cf. the work of Cohen and Levesque
PRS will perform its means-ends reasoning and 1987], who base their axiomatization of rational be-
other planning in the context of its existing inten- havior solely on this kind of mechanism).
tions. For example, consider that PRS has adopted Another good reason for dropping a goal is simply
the intention of achieving a goal g by accomplish- to fail in all attempts at achieving the goal. This re-
ing the subgoals g1, g2, and g3, in that order. In quires no more knowledge than the system acquires
the process of determining how to accomplish these in its striving for the goal, and thus is appropriate
subgoals, the system will not reconsider other means to be used as the default method for the basic inter-
of achieving g. That is, it is committed to achieving preter. But it brings us back to the former question:
g by doing g1 , g2, and g3, even if circumstances have what attempts does the system make to achieve a
so changed that there is now a better way to achieve given goal?
g than the one chosen. The gain here is in reducing In PRS, the basic system interpreter tries, exactly
decision time | in highly dynamic domains it is not once, every possible KA instance that can possi-
bly achieve the goal. It does not ask that previ- Let p be an upper bound on the execution times
ously achieved goals be reachieved (in some other of the primitive actions that the system is capable
way), nor does it try the same KA instance more of performing. Let's also assume that n is an upper
than once. In this sense, it is equivalent to a \fast- bound on the number of events that can occur in
backtrack" parser. There are good reasons for these unit time, and that the PRS interpreter takes at
choices. most time t to select the set of KAs applicable to
Unlike o-line planning or the parsing of sen- each event occurrence.2 By calculating the number
tences, PRS operates in the real world rather than of events that can occur in a single cycle of the PRS
an hypothetical one. Thus, the actions it takes can- interpreter, it is not dicult to show George and
not be withdrawn and alternative approaches tried Ingrand, 1988] that the maximum reactivity delay is
in their stead. Once some goal has been achieved, it R = p=(1 ; nt), where we assume that t < 1=n.
has really been achieved | there is no point in in- This means that, provided the number of events
voking KAs solely because they represent alternative that occur in unit time is less than 1=t, PRS will
ways to achieve an already achieved goal. notice every event that occurs and is guaranteed to
The remaining issue concerns the number of tries do so within a time interval R. (If the event rate
we make of a single KA instance to achieve a given exceeds 1=t, the system may not be able to keep
goal. It is, of course, quite possible that, where the abreast of the changes in the environment.) In the
rst try does not succeed, the next will, even if we shuttle application, which includes over 100 KAs and
carry out exactly the same actions as we did the rst 1000 facts about the RCS, the values of p and t are
time. However, to determine if retrying could suc- less than 0.1 seconds, giving a reactivity delay of at
ceed would, in most practical cases, require knowl- most 0.2 second for an event rate of 5 events per
edge of the state of the world that goes well be- second.
yond that available to the system. In the absence Because metalevel procedures are treated just like
of such knowledge, the system therefore tries each any other, they too are subject to being interrupted
applicable KA instance exactly once. (The RAP after every primitive metalevel action they take.
system Firby, 1987], on the other hand, continues Thus, reactivity is guaranteed even when the system
to reinvoke already-tried task networks clearly, this is choosing between alternative courses of action or
presents the serious problem of inde nite looping.) performing deductions of arbitrary complexity.
Once all attempts at achieving a given goal have Having reacted to some event, it is necessary for
been exhausted, it is still possible for some met- the system to respond to this event by perform-
alevel KA to respond to this failure and invoke yet ing some appropriate action. As the system can
other means to achieve the goal. For example, a be performing other tasks at the time the critical
meta-KA could invoke certain deductive machinery, event is observed, a choice has to be made concern-
or could decide to retry some KA instances that, ing the possible termination or suspension of those
although having been tried once, appear (for some tasks while the critical event is handled. Further-
reason known to the meta-KA) to be worth trying more, it may be necessary to decided among dier-
again. In this way, the provision of additional de- ent alternatives for handling the event. With appro-
cision knowledge can always override the basic pro- priate metalevel KAs, it is possible to guarantee a
cessing method described above. bound on this decision time George and Ingrand,
1988]. However, the construction of such algorithms
5 Planning in Real Time remains one of the more interesting areas of research
in the design of real-time systems Bratman et al.,
5.1 Guaranteed Reactivity 1988, Dean and Boddy, 1988].
Response time is one of the most important mea- 5.2 Planning or Not?
sures in real-time applications if events are not han- There has always been some confusion in the liter-
dled in a timely fashion, the system can go out of ature about the notion of planning, especially with
control. Yet few existing real-time systems are guar- respect to the kind of practical reasoning that PRS
anteed to respond within a bounded time interval performs. In the AI literature, planning is viewed
Laey et al., 1988]. as the generation of a sequence of actions to achieve
Response time is the time the system takes to rec- some given goal. The classical approach to this prob-
ognize and respond to an external event. Thus, a lem is to simulate the eects of performing the ac-
bound on reaction time (that is, the ability of a tions so as to ensure that their execution does indeed
system to recognize or notice changes in its envi- achieve the required goal.
ronment) is a prerequisite for providing a bound on However, this is not the only way to construct an
response time. PRS has been designed to operate eective plan. For example, the choice among al-
under a well-de ned measure of reactivity. Because ternative courses of action could be based on the
the interpreter continuously attempts to match KAs
with any newly acquired beliefs or goals, the system 2 As selection of KAs does not involve any general
is able to notice newly applicable KAs after every deduction beyond uni cation and evaluation of a boolean
primitive action it takes. expression, an upper bound does indeed exist.
expected time to complete the actions, or the like- Acknowledgments
lihood of success of the plans as gained through ex-
perience. In any case, simply making the choice as Pierre Bessiere, Amy Lansky, Anand Rao, Lorna
to which course of action to pursue, no matter how Shinkle, Joshua Singer, Mabry Tyson, and Dave
one does it, constitutes forming a plan to achieving Wilkins helped in the development of PRS and ex-
ones goals. tended the implementation as needed. The project
This is exactly the way PRS operates. The has also bene ted greatly from our discussions with
method of choosing between alternative courses of Michael Bratman, Phil Cohen, David Israel, and
action is embedded in the metalevel KAs of the sys- Martha Pollack.
tem and thus, in essence, the particular approach to References
forming plans is not hard-wired into the system. To
the extent that the choice is made arbitrarily, one Bratman et al., 1988] M. E. Bratman, D. J. Israel,
may wish to avoid calling this process \planning." and M. E. Pollack. Plans and resource-bounded
But where it is based on any information at all, no practical reasoning. Comp. Linguistics, 1988.
matter how meager, the determination of an appro- Bratman, 1987] M. Bratman. Intention, Plans, and
priate course of action is indeed a form of planning.
In the RCS example, the system decides between Practical Reason. Harvard University Press, Cam-
dierent courses of action depending on how the KA bridge, Mass, 1987.
was invoked and what sort of priority it has. This Brooks, 1985] R. A. Brooks. A robust layered con-
is clearly quite a weak form of planning, and more trol system for a mobile robot. TR 864, AI Lab-
complex meta-KAs | taking time availability, costs, oratory, Massachusetts Institute of Technology,
and bene ts into account | could be expected to Cambridge, Mass, 1985.
improve system reliability. However, it is interesting Cohen and Levesque, 1987] P. R. Cohen and H. J.
to observe just how weak the planning component Levesque. Persistence, intention, and commit-
can be when we have a wealth of experience (a rich ment. In Reasoning about Actions and Plans,
set of object-level KAs) to assist us. pages 297{340. Morgan Kaufmann, Los Altos,
Cal, 1987.
6 Conclusions Dean and Boddy, 1988] T. Dean and M. Boddy.
An analysis of time-dependent planning. In
The system described above was implemented on a Proc. Sixth National Conference on Articial In-
Symbolics 3600 Series LISP machine and has been telligence, pages 49{54, Seattle, Wash, 1987.
used to detect and recover from most of the possi- Firby, 1987] R. J. Firby. An investigation into reac-
ble malfunctions of the RCS, including sensor faults, tive planning in complex domains. In Proc. Sixth
leaking components, and regulator and jet failures. National Conference on Articial Intelligence,
This was accomplished by using multiple commu- pages 202{206, Seattle, Wash, 1987.
nicating instantiations of PRS and a simulator for George and Ingrand, 1988] M. P. George and
providing real-time input to the system. Complete
details of this large-scale application are given else- F. F. Ingrand. Research on procedural reason-
where George and Ingrand, 1988]. ing systems. Final Report, Phase 1, AI Center,
The experiment provided a severe and positive SRI International, Menlo Park, Cal, 1988.
test of the system's ability to operate pro ciently George and Lansky, 1986a] M. P. George and
in real time, to weigh alternative courses of action, A. L. Lansky. Procedural knowledge. Proc. IEEE
to coordinate its activities, and to modify its inten- Special Issue on Knowledge Representation,
tions in response to a continuously changing environ- 74:1383{1398, 1986.
ment. In addition, PRS met every criterion outlined George and Lansky, 1986b] M. P. George and
by Laey et al. 1988] for evaluating real-time rea- A. L. Lansky. A system for reasoning in dynamic
soning systems: high performance, guaranteed re- domains: Fault diagnosis on the space shuttle. TN
sponse, temporal reasoning capabilities, support for 375, AI Center, SRI International, Menlo Park,
asynchronous inputs, interrupt handling, continuous Cal, 1986.
operation, handling of noisy (possibly inaccurate)
data, and shift of focus of attention. George and Lansky, 1987] M. P. George and
The features of PRS that, we believe, contributed A. L. Lansky. Reactive reasoning and plan-
most to its success at this task were (1) its partial ning: An experiment with a mobile robot. In
planning strategy, (2) its reactivity, (3) its use of pro- Proc. Sixth National Conference on Articial In-
cedural knowledge, and (4) its metalevel (reective) telligence, Seattle, Wash, 1987.
capabilities. In particular, the manner in which the Kaelbling, 1987] L. P. Kaelbling. An architecture
system integrates its means-ends reasoning with the for intelligent reactive systems. In Reasoning
use of decision knowledge is considered an important about Actions and Plans, pages 395{410. Morgan
component of rational activity. Kaufmann, Los Altos, Cal, 1987.
Laey et al., 1988] T. J. Laey, P. A. Cox, J. L.
Schmidt, S. M. Kao, and J. Y. Read. Real-time
knowledge-based systems. AI Magazine, 9:27{45,
1988.
Wilkins, 1988] D. E. Wilkins. Practical Planning:
Extending the Classical AI Planning Paradigm.
Morgan Kaufmann, Los Altos, Cal, 1988.