B-SURE: A BELIEVED SITUATION AND 
UNCERTAIN-ACTION REPRESENTATION ENVIRONMENT 
JOI/N K. MYI';1LS 
ATR Interpreting Telephony Research l,aboratories 
Sanpeidani, lnuidani, ,qcika-cho, Soraku-gsn, l(yoto 619-02, Japalt 
myers@al.r ht.atLco.jp 
Abstract 
This paper l)reseltts a system that is c~q)abte of representing sit- 
us,loss, states, anci nondeterminlstic lIOIlinOllOtollic Oil'COllie 
actions ,lccmTillg ill multiple possible worlds. :\['he systcnl sup 
ports explicit representations of actions and situations used 
in intentional action theory and situation theory. \[lath types 
mid instances ere supported. Situations ¢md statc"a before a,d 
aftel-llOllll/OllOtOnic actions c~ul be repl'esellted shnultaneously. 
Agents have free will as to whether to choose to peiform an ac:- 
lion or not. Situations itud actions can have expected values, 
allowing the system to support decision making anti decision- 
based pleal isfferencing. The system cau perform global rea- 
soning simultaneously across multiple possible worlds, without 
being forced to extend each world explicitly. The resulting 
system is useful for retch *tatural language t~-~ks a~ plan recog- 
nition, intentions modeling, attd parallel ta.~k scheduling. 
1. Introduction 
The key to good reasoning is a powerful repre- 
sentation system that is able to accuratcly model 
details of a problem. Once a good represent.at,on 
has been established, problem computations often be- 
come straightforward. 
I¢~ecent advances in situation theory \[BP83,Bax89\] 
and the theory of intentions \[Bra87\] have offered 
ninny new insights on significant problems found in 
natural-language understanding. However, these the- 
aries offer philosophical approaches only, and do not 
give instructions for building concrete reprcsentation 
and reasoning engines. 
At the same time, the software systems that have 
been built for reasoning and representation fall short 
in any ruunber of areas. Production systems and se- 
mantic networks can follow chains of inferences, but 
can only represent one possible world at a time 
they cannot reason with states that are both pos- 
sibly true and possibly not true, while keeping the 
chains of resulting inferences separatc. Most plasl- 
nets work with limited possible worlds, but callnot 
reason and perform inferences across multiple worlds 
at the salne time. The classical ATMS 1 call represent 
and reason with multiple timeless possible worlds, but 
calmot represent actions \[dK86\]-in particular, non- 
monotonic actions where a retracted state is both 
believed'to be truc in the world before ttte action 
takes place, and believed to be not true in the world 
representing the situation after the retracting action 
has taken place, camtot be represented. In addition, 
the ATMS only represents propositions that are in- 
stant'steal constants or Skolem constants; it does not 
represent uninstantiated variables. A modified ATMS 
that can represent nonmonotonic transitions between 
worlds has been developed \[MN86\], but this system 
does not explicitly represent situation types and in- 
stances, action events, nor nondeterminism. Most 
plan inference systems have ignored free will and the 
1Aanumption-13tmed Truth Maintenta*ce Systenl \[dK86\] 
explicit representation of the right to choose actions, 
e.g. to choose to he uncooperative. Almost all prcvi- 
ous systems |lave ignored the nondetermimstie qual- 
ity of real-world actions that aecessitatcs commit- 
meat ill intentions. Real actions call result in one 
of several possible outcomc situations, where,xs al- 
olost all previous systems are are completely unable 
to model uondcterminlstie outcomes. Only dccision- 
analysis systems have modeled cxpected wdues of 
actions, alnl they do not support inferencing. See 
\[BL85\] fi)r an excellent summary of issues. 
Tile B-SURI~ (Believed Situation and Uncertain- 
action Representation Environment) packagc is an 
implemented system that supports representation, 
phmnmg, decision-making, and,plan recognition us- 
ing probabilistic ,and uncertain actions with non 
detcrministic outcomes in multiple possiblc action 
worlds. Situations, states, and action events ,axe all 
represented explicitly, using types (wtriables) and in- 
stances. The B-SURE systcm is iml)lemented as a se- 
ries of extensions to a classical ATMS. The resulting 
system is very useful, and is being used in plan recog- 
nition, intentional agent, and scheduling research. 2. Situation Theory 
In \[BP83\], situations are divided into the categories 
abstract and real, and also into the categories "states 
of affairs" alid "courses of events". Abstract situ- 
ations denote situations that are mental representa- 
tions. All the situations discussed in this paper are 
"abstract situations". Real sitnations denote situa- 
tions as they actually are in the real world. Since it 
basically never makes sense to talk about real situ- 
ations in tile computer, there is no need to snpply 
these in a representation environment. "States of af- 
fairs" correspond to situations that axe static, called 
simply situations in this paper. "Courses of events" 
correspond to situations that describe actions that 
are being executed, called action events or actions in 
this paper. Barwise and Perry also make use of "rela- 
tions" defined over '*individuals" and "space-time lo- 
cations". This paper takes as primitive the expression 
of a relation, which will be termed a stats. The user is 
free to mention individuals or space-time locations in 
state descriptions as desired. State descriptions may 
be represented nsing logical forms, feature structures, 
or other methods~sinee the contents of states are not 
used by B-SURE except for output, it does not mat- 
ter. States, situations, and actions axe assigned one 
of the belief values {definitely believed tzate, 
possibly believed true, not believed true, 
believed not true, not believed}, other- 
wise known ms {actua/, possible, bypotheticed.. 
inconsistent, null}, corresponding to the amount 
of snpport off, red by tile system's underlying ATMS 
ACRES DE COLING-92, NAbffES, 7-3-28 AO(n" 1992 9 6 i PRec. OF COLING-92, NANTES, AUG. 23-28, 1992 
precondition \] ~ | outcome i_\[-In2ge. • , +,,.+~,,,. +,,.. ~ ~cutm !~ trsnsition -*'I ---~ situation tripe ~ o( tUpes 
I ........... ,q~ ~e- JI t~___JI tuoe I~~ t ~-- i ~ pus 
l- PI~ '"trensiti0n #n rill ~ ~ state 
I~~~st~o~e ~ /~°!-~et~-\]YP °sitiv° 
I action #n) -- ~ j ~obtcome I 
v~ "--'~'~-~ ~.~ \[situetioninstence *ll KEY happens n • 
ATMS implies ~ II 
ATMS mutualI V-~I 
J -exclusive ~=sumpt, ons)l 
Figure 1: Structure for Representing Nondeterministie Actions 
~(Possible Outcune 51tnation':#i) (5tart| ng 51 tuatlon~ ....... 
~ "~'-"~(...Possible odtcone =51tuation, In ~) 
Figure 2: Compact Graphical Representation which Omits States and Types 
representation (see \[Mye89\] for more information). 
~. Intentional Action Theory 
One model of intentions states that an intention is 
a choice to perform an action, plus a commitment to 
obtaining its desired outcome\[CL87\]. With determin- 
istic action outcomes, there is no real need for endeav- 
oring \[Brag7\], since once the action has been started, 
it is guaranteed to finish properly. Many planners in 
fact operate in this "fire and forget" mode. However, 
once it is acknowledged that action execution is in fact 
nondeterministic and can have undesirable outcomes, 
the need for endeavoring becomes clear. The planner 
must predict the likelihood of possible outcomes hap- 
pening, and judge which action sequence offers the 
best chances. It must interactively maintain a his- 
tory of past endeavors and results, and modify its 
future behavior based on current outcomes. Acting 
intentionally becomes significantly more interesting 
and realistic with the explicit representation of pos- 
sible chains of nondeterministic actions. 
4. Previous Efforts 
DeKleer \[dK86\] presents the first ATMS. Morris 
and Nado \[MN86\] present an ATMS that can represent 
nonmonotonic transitions, but do not handle proba- 
bilities, uncertainties, explicit situation types, state 
types, nor action events. Tile research of Allen (e.g. 
\[AK83, A1187\]), who uses a predicate-calculus rep- 
resentation, offers some of the best multiple-worlds 
(deterministic) action representation in this field. 
Charniak and Goldman \[CG89\] use probabilities and 
Bayesian nets to represcnt the truth value of prob- 
abilistic statements and attack story understanding. 
Although nondetcrministic-outcomc actions are not 
represented, and Bayesian nets cannot support global 
inferencing with nonnronotonic actions, their work is 
important. Norvig and Wilensky \[NW90\] comment 
on problems of probabilistic statements. "1'he most 
similar work is recent research by Rao and Georgeff 
(e.g. \[RGgl\]), who use a modal logic instead of an 
ATMS to represent nondeterministic actions. 
5. B-SURE Entities & Implementation 
The underlying ATMS works with nodes, assnmp- 
tions, and implications (justifications). See \[dK86\]. 
A slate consists of a proposition about the world. 
States are primitives. A situation is a set of positive 
and negative (withdrawn) states. An action event 
represents the state that "execution of the action has 
started". States, situations, and actions have types 
and instances. See figure 1. (The abridged repre- 
sentation of figure 1 is shown in figure 2.) Existance 
of an instance in a world always implies existance of 
its type. A chooses node is an assumption associ- 
ated with an action instance that represents whether 
an agent chooses to execute that action or not. The 
chooses assumption together with the starting situ- 
ation instance imply the action instance. Since an 
agent typically can only execute one action in a given 
situation, the situation's ensuing chooses assumptions 
are rendered mutually exclusive (pairwise "nogood"). 
Action types have precondition situation types. Ac- 
tion instances are instantiated from types by first ver- 
ifying that the precondition situation type is believed 
true in that world. Action instances transition from 
a starting situation instance to ouc of a number of 
known nondeterministic outcome situation instances. 
Actions have transitions. A transition has an out- 
con,e situation and a probability or an uncertainty. 
An uncerlaiuty is defined as a probability random 
variable of range I0, 1\] together with an associated 
second-order probability distrilmtion. Uncertainties 
are initialized using maximum-entropy theory, and 
get updated as outcome observations are taken, to en- 
able the system to learn and estimate possible proba- 
bilities. See Section 6. Uncertainties are used to rep- 
resent confidence in likelihood values and to make de- 
cisions regarding information-gathering activity. The 
calculus of uncertainties is too complex to explore fur- 
ther here, and is not required for understanding tile 
mum capabilities of the representation; probabilities 
are sufficient. Transitions can be types or instances. 
ACRES DE COLING-92, NANIT~. 23-28 AOt~rr 1992 9 6 2 PROC. OF COL1NG-92, NANTES, AUO. 23-28. 1992 
A transition instance is defined as a happens mssump- 
tion. An action instance, together with a happens 
assumption, iulply the corresponding outcome situa- 
tion instance. Typically only one outcome situation 
can occur from a giveu action instance, so the action's 
happens assun-|ptions are nlade mutually exclusive. A 
situation type is implied by its state types. When an 
outcome situation instance is iastantiated, all of its 
new positive states are instantiated and all of its ohl 
negative states are retracted. A positive nonperma- 
nent state instance is implied by a nol-relracled-yel 
assumption. The outcome situation instance remem- 
bers these. Situation and action instances store an ex- 
plicit environment history of all added state, chooses, 
and happens assumptions that are currently believed 
true in that possible world's timelinc. A negative 
state is retracted by making the situation instance 
and the state's "not-rctrasted yet" ,assumption mutu- 
ally inconsistent, and deleting the state's assumption 
from the outcome situation's environment history. A 
state type or instance or situation type's belief value 
in a particular world is found by testing that node 
against a situation instance's environment history. 
Situation types and instances can have values. Ac- 
tions can have costs. The expected value of an action 
is determined by summing the transition probabilities 
times the expected values of the outcome situations, 
when known, and subtracting its cost. Tim expected 
value of a nonvalued situation instance is determined 
by maximizing the expected values of the possible 
subsequent actions, when known. In this manner, 
decision theory determines the course of action with 
the maximum expected value at any one situation, 
for a planning agent. This can bc used to predict the 
probable next course of action of a planning agent by 
an observing agent performing plan recognition (ac- 
tually, "decision recognition") \[Mye91\]. 
6. Probability Estimation 
The probability of an outcome situation i occur- 
ring following performance of an uncertain action is 
estimated using the new estimator ~ instead of ~, 
where m is the total number of previously observed 
trials of that action type, k~ is the previously observed 
number of ith situation-type outcomes, and n is the 
number of known possible outcome situations from 
that action. The new estimator is optimal. It repre- 
sents the center of mass of all possible probabilities, 
instead of the maximum-likelihood mode; it converges 
faster and on average is more accurate than the old 
estimator; and, it can be used accurately with small 
sample numbers and small snccess counts \[Mye92\]. 
7.Maintaining an Interactive History 
One important advantage of the B-SURE system is 
that not only can it be used for hypothetical reason- 
ing about future events, but the same structures can 
then be used as a history mechanism for interastively 
monitoring and representing the history of the ac- 
tual events as they occur. A user system should start 
out in a known situation, which is presumed actual. 
Typically, the user system will use B-SUrtE to explore 
many different nondeteeministic-action sequences and 
make decisions ~s to which actions are tile best ones 
to perform. The system will then start executing the 
first action m tile chosen sequence. At. this point, 
tile user system should instruct tile B-SUItE system 
to presume the chooses assumption associated with 
tile chosen action being executed, which will change 
its truth value from "possibly believed true" to "deli- 
nitely believed true". If the chooses node has already 
been made inconsistent with other chooses nodes (be- 
cause the user-system or agent could ouly perform one 
action at a time), those other nodes are automati- 
cally rendered "believed not-true" at this point. The 
presumption of the chooses node renders the associ- 
ated implied Action Event instantiation "definitely 
believed true" at this point, ,also. This represents the 
fact that the action has stated and is currently being 
execnted. 
When the action finishes, it is necessary for the 
B-SURE system to realize which outcome occurred. 
This is typically performed by the system setting up a 
recognition demon that is attached to a separate state 
or situation type that, when true, reliably indicates 
that a given outcome has occurred. When the demon 
fires, it presumes the outeome's happens assumption. 
It is important to ensure ttlat one and only one recog- 
nition demon fires. Alternatively, tile user can control 
presuming the happens nodes directly. When a sin- 
gle happens assumption is presumed, it automatically 
renders its sibling happens assumptions "de~init ely 
believed not;-tzale". 
The combination of tile happens node being pre- 
sumed and the action event node already being be- 
lieved true renders the appropriate resulting situation 
instance believed true. Note that if any instance be- 
comes true, so does its associated type node as well. 
At any one point in tinm, the states, situations, 
and action event instances that have happened in the 
world already are believed true; and the situations 
and events that have not happened yet but could 
happen are believed possible. In this way, the sys- 
tem maintains a timehne history of the situations and 
action events that have in fact occurred, while allow- 
ing hypothetical planning and exploration of possible 
future events in the same data structure. 
It is not necessary for the system to maintain only 
a single timeline history. It is possible to maintain 
disjoint histories, to represent e.g. progress made by 
different processing agents, progress made in different 
domains, or progress made at different hierarchical 
levels of abstraction. It is possible to maintain forking 
(nondisjoint) histories if this makes sense, and tim 
mutual exclusion options have been turned off (see 
Section 10). 
Counterfaetuals The system maintains the 
structures of past possibilities that did not happen. 
Although these are not believed true, it is possible 
for the user to explore these structures and perform 
reasoning on what could have occurred had certain 
actions been chosen or certain nondeterministic out- 
comes happened, by supplying an extra counierfac- 
tual assumption to justify the desired action or sit- 
ACRES DE COLING-92, NANTEs, 23-28 ^o13"r 1992 9 6 3 Paoc. OF COLING-92, NAI, rrEs. AUG. 23-28, 1992 
O ere |I\[~et\[r~g~ /(tnKe-suscsuEtt~~ 
A \[RECP OFFICE\] \[x~r ,rEd~cTe . / ................. ~(flEEOS-RIOE CfiLLEr)~/I \[CONT RIuE-rIOEIIj X ......... OF~CE~)~ ~(Geta there Late) 
Gets there Late ~(TAKE-TAXI 
CALLER)~ ---- -- ~ ' ~(G~ts there On T~ 
Figure 3: Modeling a Plan/Decision Inference Problem in Getting to a Conference On Time 
uation instance. It is even possible to add to these 
structures, if necessary. This can be used to explic- 
itly represent newly-received p~st connterfactual iu- 
formation (e.g., "If you had applied for the conference 
last June, the cost would have been 35,000 yen") and 
the associated reasoning derived from snch assertions. 
Such reasoning has traditionally been very difficult to 
represent, because of the negative truth values. 
8. Decision Inference Example 
A researcher is calling a conference office from tbe 
tram station and wants to get to the conference on 
time. He has a choice between asking for taxi di- 
rections, or requesting the office to send a shuttle- 
bus out directly to give him a ride. The shuttle will 
take him directly to the conference on time. If he re- 
quests and the office turns him down, he has a choice 
between taking a taxi, and taking the regular bus. 
These cost ditferent amounts of money and have dif- 
ferent chances of getting to the conference on time. 
See figure 3. The plan inference system must pre- 
dict which paths of information he will explore, i.e. 
what he will say next; and then which decisions he 
will make for his actions. This is done using "deci- 
sion infcrence", by understanding which action trees 
offer the best expected value based on the value and 
chances of outcomes. Note that the shuttle-bus, the 
taxi, and the regular bus will all three allow the re- 
searcher to possibly obtain his desired goal, but there 
are definite preferences. The system should not re- 
main uncommitted. See \[Mye91\] for more details. 
9.Intentional Communication Example 
A recent analysis of 12 actual interpreted telephone 
conversations revealed that 31% of the utterances 
were spent in requests for confirmation and repeti- 
tions of information such as telephone numbers, name 
spellings, and addresses, that were not completely un- 
derstood the first time \[OCP90\]. This means that the 
traditional plan-recognition model of assuming that 
the hearer automatically understands the semantic 
content of the speaker's utterance is fallacious. The 
speaker, and the system too, must consider the case 
in which the hearer does not understand an utter- 
ance. Since the speaker wants and intends to commu- 
nicate specific information 2, the speaker will endeavor 
to ensure that the information is communicated, by 
repeating an utterance when it is not understood. 
Thus, speaking an utterance is a nondeterministie ac- 
Lion; it, is unclear whether the hearer will uuderstaald 
or not. Intentional utterauce acts are therefore mod- 
eled ~ nondeterministic-outcome actions by B-SURE. 
l)ifi'erent courses of the conversation cat\] be repre- 
sented depending upon the outcomes of the utterance 
acts. See Figure 4. 
10. Process Scheduling Example 
The application of the BEHOLDEIL a limited-resource 
parallel sehednling system to translation systems is 
being researched. A hypothetical model system is 
used for testing. The system will accept an input 
caa|didate from a speech recognition module, and at- 
tempt to quickly transfer tim result directly to out- 
put. If required, a morphologicd analyzer will derive 
multiple possible analyses candidates for each input 
candidate. A pattern marcher will then recursively 
apply a body of patterns to each analysis candidate. 
Each pattern has a series of transfer-driven transla- 
tion templates; each template has a series of prototyp- 
icai exmnple bindings. The highest-ranking structure 
of matching nested patterns and their bindings are 
sent to a template marcher. The distances between 
the pattern bindings and the template examples for 
each pattern in the structure are compared using a 
thesaurus. The template with the closest match for 
each pattern will be used to assemble a translation. 
It is the responsibility of the BEHOLDER system to 
schedule this activity in an opportunistic fashion on 
multiple processors. There is no need to continue to 
explore a branch if a good translation has been found. 
The BEH OLDER system must use value-of-information 
theory and decision theory to determine which pro- 
cess branches to explore next and when to stop. 
The BEIIOLDER scheduler uses the B-SURE system 
to keep track of which processes are running and 
which have been executed. Using this representation, 
it can plan ahead and decide how useful it is to ex- 
pand a particular path of execution. As processes are 
started, the chooses nodes are presumed. Figure 5 
shows a simulated run where the direct transfer, the 
morphological analysis, one pattern match, and one 
template match have been run. The template match 
has examined two examples so far. 
Since in this ease more than one a~tion can be exe- 
cuted at a time, and one action cau legally have more 
than one possible outcome, it was necessary to mod- 
ify the n-SORE system to allow local disabling of the 
2Note that people do not always decide to intend to en- 
deavor to do everything that they weatt. Intending is qlfite 
different from wmlting. 
aBeneficlal Entity for Heuristically Ordering processes un- 
der Limited resources and Decision-making for Execution in 
Real-time. 
ACRES DE COLING-92, NANTES, 23-28 AO(rr 1992 9 6 4 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
Doesn't Understand) ~Ir~~~ neuter 
(~Oea er ntens ,/ ...... ~Hearer Understands ) 
k to Say "lZJ-qSfiT" ~ ...... yIT~ ~ J~ ~ -(H ..... D .... "t Understand) 
~lea~~~ Understands) 
Figure .1: Mode}rag an Intention to Conrmuuicate a Telephoto. Number Correctly 
fNo Good Resuits Obtk|ned • |DIrEcT_TR/~NsFErkI~ 
,, ,. ~, 
.oRPh-AN~LY5 \[-- 
l"igore 5: S~ 
mutually-exclusive actions and outputs features. \[CL87\] 
11. Conclusion A powerful situation represen- 
tation tool is required for representing past, present, 
and future nonmonotonic actions, when the actions \[dK86\] 
can have nondcterministic outeonms. The B-SURE 
euviromnent offers such a tool. Being able to model 
realistic actions allows exploration of significant prob- \[MN86\] 
|eros in situation modeling, plan inference, intentional 
actions research, and value-of-information theory 0.~ 
applied to parallel process scheduling. \[Mye89\] 
References 
\[AK83\] James F. Allen and Johannes A. Koomen. Plan- 
ning using a temporal world model, ht IJ- \[Mye91\] 
CAI'83, pages 741-747, Kaxlsruhe, 1983. 
\[Al187\] James Allen. Natural Language Understanding. 
Benjamin/Cummings, Menlo Park, CA, 1987. \[Mye92\] 
\[Bar89\] Jolt Barwise. The Situation in Logic. Center for 
the Study of Language and Information (CSLI), 
Stauford, CA., 1989. 
\[BL85\] lLonald J. Brachman and Hector J. Levesque. 
lleadi*Jgs it~ Knowledge Representation. Morga~t \[NW90\] 
Kaufl~tann, Los Altos, CA, 1985. 
\[BP83\] Jon Barwise and John Perry. Situations and 
Attitudes. The MI'F Press, Cambridge, 1983. 
\[Bra87\] Michael E. Bratman. Intention, Plans, and \[OCP90\] 
Practical Reason. Harvard Univ. Press, Cam- 
bridge, MA, 1987. 
\[CG89\] Eugene Charniak and Robert Goldman. A \[RG91\] 
semanti~ for probabilistic quantilier-free first- 
order languages, with particular application to 
story understanding. Ira 1JG'Al'89, pages 1074 
1079, Detroit, MI, 1989. 
Pit}lip R. Cohen and Hector J. Levesque. In- 
tention : choice + commitment. In AAAl'87, 
pages 410-415, Seattle, WA, 1987. 
Johtm de Klcer. An Assumption-based TMS. 
Artificial b~telligence, 28(2):127-162, March 
1986. 
Paul I1. Morris and ttobert A. Nado. Represent- 
ing actions with an Assmnption-based TMS. In 
AAAl'86, Philadelphia, PA, 1986. 
John K. Myers. An a-ssumption-bmsed plan in- 
ference system for conversation understanding. 
In WGNL Meeting o.\[ the \[PSJ, pages 73-80, 
Okinawa, Japan, June 1989. 
John K. Myers. Plan inference with 
probabilistic-outcome actions. In Conj. Proc. 
lnJormation Processing Society of Japan, vol- 
ume 3, pages 168-169, Tokyo, March 1991. 
John K. Myers. An introduction to plannIng 
and meta-decision-making witt~ uncertain non- 
deterministic actions using 2nd-order probabili- 
ties. In First Inter. Conferettce on AI Planning 
Systems, College Park, Maryland, June 1992. 
Peter Norvig and Robert Wilensky. Ab~ 
duct}on models for semantic interpretation. 
In G'OLING-90, volume 3, pages 225-230, 
Ilelsinki, Finland, August 1990. 
Sharon L. Oviatt, Philip R. Cohen, and Ann 
Podlozny. Spoken language in interpreted tele- 
phone dialogues. TechnicM Report AI0-496, 
SRI International, Menlo Park, CA, 1990. 
Anad S. Rao and Michael P. Georgeff. Asymme- 
try thesis and side-effect problems in linear-time 
and branching-time intention logics. In Procee& 
ings of 1JCAI-91, Sydney, Australia, 1991. 
ACRES DZ COLING-92, N^tcr~s, 23 28 Ao~rr 1992 9 6 5 Proc. OF COLING-92. NANTES. AUG. 23-28, 1992 
