Zhiwen Wu and Ahmed Y. Tawfik
School of Computer Science, University of Windsor,401 Sunset Ave.,Windsor,Ontario N9B 3P4, Canada
Keywords: Chance discovery, knowledge base, relevance, planning, ontology.
Abstract: This paper argues that chances (risks or opportunities)
can be discovered from our daily observations and
background knowledge. A person can easily identify chances in a news article. In doing so, the person
combines the new information in the article with some background knowledge. Hence, we develop a
deductive system to discover relative chances of particular chance seekers. This paper proposes a chance
discovery system that uses a general purpose knowledge base and specialised reasoning algorithms.
According to Ohsawa and McBurney (2003), a
chance is a piece of information about an event or a
situation with significant impact on decision-making
of humans, agents, and robots. A ‘chance’ is also a
suitable time or occasion to do something. A chance
may be either positive –an opportunity or negative –
a risk. For example, predicting a looming earthquake
represents a “chance discovery”.
Many approaches have been applied to chance
Rare events may represent chances
known to co-occur with important events, while the
important events can be extracted using data mining
techniques. KeyGraph, the application of this
technique, was applied to various data, such as
earthquake sequences, web pages, documents
(Ohsawa et al., 1998; Ohsawa and Yachida, 1999;
Ohsawa, 2003a; Ohsawa, 2003b). Tawfik (2004)
proposes that chance discovery represents a dilemma
for inductive reasoning. Induction assumes that
current trends will carry into the future thus favoring
temporal uniformity over change. However, current
observations may lead to different possible futures in
a branching time model. Finding a proper
knowledge representation to represent all these
possible futures is important. Otherwise some
chances will be missed. Bayesian and game theoretic
approaches are presented as viable chance discovery
techniques. Abe (2003a, 2003b) considers chances
as unknown hypotheses. Therefore, a combination of
abductive and analogical reasoning can be applied to
generate such knowledge and chances can be
discovered as an extension of hypothetical
reasoning. McBurney and Parson (2003) present an
argumentation-based framework for chance
discovery in domains that have multi agents. Each
agent has a partial view of the problem and may
have insufficient knowledge to prove particular
hypotheses individually. By defining locutions and
rules for dialogues, new information and chances
can be discovered in the course of a conversation.
In this paper, we incorporate some new elements
to the chance discovery process. These elements
have implications to both the conception and
discovery of chances and can be summarized as
are not necessarily unknown
hypotheses. Many chances result from known
events and rules. For example, applying for the
right job at the right time represents a chance
for an employment seeker as well as the
employer. In this case, the goal is clear.
However, chance discovery means that the
employment seeker applies at the proper time
and for the employer, it means to correctly
project which applicant will be better for the
herently, chance discovery has a temporal
reasoning component. New risks and
opportunities are typically associated with
change. An invention, a new legislation, or a
change in weather patterns may result in many
chances. Incorporating chance discovery in a
belief update process is fundamental to this
work. Chances are relative; someone’s trash
may be another’s treasure. For example, finding
a cure for a fatal disease represents more of a
chance to an individual suffering from this
condition or at risk to contact it.
Wu Z. and Y. Tawfik A. (2005).
In Proceedings of the Seventh International Conference on Enterprise Information Systems, pages 111-118
DOI: 10.5220/0002539601110118
To discover chances and take advantage of
them, a system which can perform deductive
reasoning is needed.
Therefore, we consider chance discovery as a
process that tries to identify possibly important
consequences of change with respect to a particular
person or organization at a particular time. For this
to happen, a logical reasoning system that
continuously updates its knowledge base, including
its private model of chance seekers (CS) is
necessary. A chance discovery process may act as an
advisor who asks relevant “what if” question in
response to a change and present significant
consequences much like seasoned parents advise
their children. Such advice incorporates knowledge
about the chance seekers, their capabilities, and
preferences along with knowledge about the world
and how it changes.
In a word, to discover chances, we need three
things: First, a knowledgeable KB which can infer
and understand commonsense knowledge and that
can incorporate a model of the chance seeker.
Second, we need a source for information about
change in the world. Third, we need a temporal
projection system that would combine information
about change with the background knowledge and
that would assess the magnitude of the change with
respect to the knowledge seeker. Cyc knowledge
base is supposed to become the world's largest and
most complete general knowledge base and
commonsense reasoning engine and therefore
represents a good candidate as a source for
background knowledge. Information about changes
occurring in the world is usually documented in
natural languages. For example, a newspaper can
serve as a source for information about change. We
need Nature Language Processing (NLP) tool to
understand this newspaper. We assume that Cyc
natural language module will be able to generate a
working logic representation of new information in
the newspaper. However, for the purpose of the
present work, understanding news and converting it
to Cyc representation has been done manually. This
paper proposes an approach for assessing the
implications of change to the chance seeker and
bringing to the attention of the chance seeker
significant risks or opportunities.
The paper is organized as follows: Section 2
establishes the notion that chance and change are
tied together. Section 3 introduces Cyc knowledge
base and its technology. Section 4 presents the
chance discovery system based on Cyc.
Chances and changes exist everywhere in our daily
life. In general, changes are partially observable by a
small subset of agents. Therefore, it is more likely to
learn about changes happening in the world through
others. For example, information about change could
be deduced from conversations in chat rooms,
newspapers, e-mail, news on the WWW, TV
programs, new books and magazines, etc. In another
word, change causing events occur daily around the
world. The amount and rate of those events is very
large. However, a relatively small portion of these
changes represent risks or opportunities to any
particular chance seeker.
Initially, the system starts with a stable
knowledge base KB. The knowledge base represents
the set of widely held knowledge. As part of KB’s
knowledge, each chance seeker maintains its own
private knowledge that describes its current
attributes. In addition to KB, each chance seeker
also maintains its private goals and plans about how
to achieve those goals. If chance seeker doesn’t
maintain its goals, the system will use default goals
that are widely accepted as common goals. For
example, the system assumes that all people want to
become more famous or richer, want their family
members and relatives to be rich and healthy, etc.
We assume that the chance seeker has already
exploited the chances present in the current KB and
that the current plans of chance seeker are the best
according to current KB. However, current plans
may only be able to achieve part of the goals. For
example, the goal to own a house in Mars is
unachieved by current knowledge.
A goal of chance seeker can be represented by a
set of sentences describing a future status of chance
seeker’s attributes. For example, if chance seeker set
up the goal to be a famous scientist, the system can
judge the achievement of the goal by measuring
chance seeker’s current attributes, such as education,
occupation, published papers, social class, etc. The
system maintains an attribute framework of chance
seeker in KB. The attribute framework can be able
to change if necessary. A goal can be considered as a
future projection of current framework. On the other
hand, a future set of attributes could satisfy many
goals of chance seeker. Current plans of chance
seeker project current set of attributes to the most
achievable set of attributes.
As new information B becomes available, an
update operation is triggered. The update operation
proceeds in two phases: a explanation phase and a
projection phase. The explanation phase tries to
revise current plans that may have been proven to be
inaccurate by the occurrence of B. Similarly, the
projection phase, revises current plans to take into
account the occurrence of B. A risk is detected if the
occurrence of B results in a threat to the causal
support for one of the plans of the chance seeker. An
opportunity is detected if B satisfies one of the
followings: the occurrence of B enables another one
of the goals of the chance seeker to become
achievable, or better plans can come up after B. In
some cases, a particular piece of new information
will result in both risks and opportunities.
The Cyc knowledge base (KB) (, 2002)
is a formal system that represents of a vast quantity
of fundamental human knowledge: facts, rules of
thumb, and heuristics for reasoning about objects
and events of everyday life. The medium of
representation is the formal language known as
CycL. CycL is essentially an augmentation of first-
order predicate calculus (FOPC), with extensions to
handle equality, default reasoning, skolemization,
and some second-order features. For example:
(#$forAll ?PERSON1
(#$isa ?PERSON1 #$Person)
(#$thereExists ?PERSON2
(#$isa ?PERSON2 #$Person)
(#$loves ?PERSON1 ?PERSON2))),
in English, means
“Everybody loves somebody.”
In Cyc, a collection means a group or class.
Collections have instances. Each instance represents
an individual. For examples,
(#$isa #$AbrahamLincoln, #$Person).
(#$isa #$BillGates, #$Person).
Abraham Lincoln and Bill Gates are individuals.
Person is a collection. A collection could be an
instance of another collection. For example,
(#$genls #$Dog, #$Mammal),
means “Collection Dog is an instance collection
of collection Mammal”.
In other word, Dog is a specialization of
Mammal. It can be said that every individual is an
instance of Thing, which is the most general
collection in Cyc KB. Some individuals could be
part of other individuals. For example, Microsoft is
an individual. Joe works for Microsoft. Joe is part of
Constants are the "vocabulary words" of the Cyc
KB, standing for something or concept in the world
that many people could know about. For example,
#$isa, #$Person and #$BillGates are constants.
The assertion is the fundamental unit of
knowledge in the Cyc KB. Every assertion consists
an expression in CycL language that makes
some declarative statement about the world
a truth value which indicates the assertion’s
degree of truth. There are five possible truth
values, including monotonically true, default
true, unknown, default false and monotonically
A microtheory of which the assertion is part of a
theory. Section 3.1 gives a detailed explanation
of microtheories.
A direction which determines whether
inferences involving the assertion are done at
assert time or at ask time. There are three
possible values for direction: forward
(inferences done at assert time), backward
(inferences done at ask time), and code
(assertion not used in regular inference).
A justification which is the argument or set of
arguments supporting the assertion's having a
particular truth value.
An assertion could be a rule or a Ground Atomic
Formula (GAF). A rule is any CycL formula which
begins with #$implies. A GAF is a CycL formula of
the form, (predicate arg1 [arg2 ...argn]), where the
arguments are not variables.
In Cyc, time is part of the upper ontology. It is a
physical quantity. A temporal object such as an
event, a process, or any physical object has a
temporal extent. The time model is interval-based
with suport for points. TimeInterval has dates, years,
and so on, as its subcategories. An event is a set of
assertions that describe a dynamic situation in which
the state of the world changes. An event has non-
empty space and time components. It may also have
performer, beneficiaries, or victims. A script in
CycL is a type of complex event with temporally-
ordered sub-events. Applications can use script
recognition – that allows them to identify a larger
script from some stated events that are constituent
parts of the script. Scripts can also be used for
planning and for reading comprehension.
3.1 Microtheories
A microtheory (Mt) is a bundle of assertions. The
bundle of assertions may be grouped based on
shared assumptions, common topic (geography,
football, etc), or source (CIA world fact book 1997,
USA today, etc). The assertions within a Mt must be
mutually consistent. Assertions in different Mts may
be inconsistent. For example,
MT1: Mandela is President of South
MT2: Mandela is a political prisoner
Microtheories are a good way to cope with
global inconsistence in the KB, providing a natural
way to represent things like different points of
views, or the change of scientific theories over time.
Mts are one way of indexing all the assertions in
Cyc KB.
There are two special Mts, one is #$BaseKB
(always visible to all other Mts), the other one is
#$EverythingPSC (all other Mts are visible to this
Mt). #$EverythingPSC is a microtheory which has
no logically consistent meaning but has a practical
utility just because it is able to see the assertions in
every microtheory.
The Cyc KB is the repository of Cyc's
knowledge. It consists of constants and assertions
involving those constants. It could be regarded as a
sea of assertions, see figure 1. Form ontology point
of view, the Cyc KB could also be thought of as
ure 1: C
c Knowled
e Base as a sea of Assertions
Figure 2: Chance Discovery Syste
made up of layers ordered by degree of generality.
Cyc uses two rules of inference in theorem proving,
modus ponens and modus tollens.
Cyc-NL is the natural language processing
system associated with the Cyc KB. It could
translate natural language into CycL. Cyc-NL has
three main components: a lexicon, a syntactic parser
and a semantic interpreter. The lexicon along with a
generative morphology component generates part-
of-speech assignments for words in a sentence. The
syntactic parser uses a grammar to generate all valid
parses for the sentence. The semantic interpreter
produces pure CycL equivalent for the input
Figure 2 shows the framework of chance discovery
system. Nature Language Processing (NLP) modules
analyze daily news and generate new knowledge
which is represented in logic. The new knowledge is
then integrated into public Cyc KB servers. The
private Cyc KB server owned by the chance seeker
will connect to public KB servers and update its
knowledge. On the other hand, the chance seeker
updates its private attributes in the private Cyc KB.
The knowledge about chance seeker can be regarded
as a virtual chance seeker living in Cyc KB. A
chance seeker sets up its goals or uses default goals
in the Goals & Plans Module. New knowledge
triggers the CD modules that measure the relevance
of the new knowledge to the chance seeker. The new
knowledge is considered to be a chance candidate if
the relevance score is above a certain threshold. By
trying to revise current plans using the new
knowledge, the magnitude of this chance candidate
can be measured using a utility evaluation process.
When the magnitude of the utility is above a
specified threshold, a chance is detected. Finally, the
system visualizes the chances to chance seeker, and
revises current plans for future chance detections.
4.1 The Relevance of New Knowledge
New knowledge is relevant to the chance seeker if it
has an immediate impact on the seeker’s attributes
or on the achievability of the chance seeker’s goals.
For example, the new knowledge that shows that the
chance seeker inherited a fortune is relevant as it
changes the seeker’s wealth attribute. The new
information can affect the achievability of goals in
three ways:
making new goals achievable,
making some previously achievable goals
unattainable, or
changing the cost or reward of achieving some
A goal is considered achievable if the system
finds a plan to the goal from the current state. To
impact the achievability of a plan, the new
knowledge could affect the causal support for
actions in the plan or the likelihood of success.
Testing the relevance of new information to the
chance seeker is desirable to filter out irrelevant
information. Fully testing the relevance of new
information with respect to its impact on the chance
seeker’s attributes and plans could be
computationally expensive. Therefore, we gradually
apply a series of relevance tests with increasing
computational cost. These tests are:
testing if the new information is subsumed by
existing knowledge,
testing for temporal relevance,
testing for spatial relevance,
testing for impact on the chance seeker’s
attributes, and
testing for impact on the chance seeker’s plans.
To verify that the new information is actually
new, and is not subsumed by knowledge already in
the KB, we test if it is entailed by existing
knowledge. For example, if the KB contains
assertions indicating that Paul Martin is the leader of
the Liberal Party, that the Liberals won the largest
number of seats in the parliament and that the leader
of the party that wins the most seats becomes the
Prime Minister. It becomes redundant to add an
assertion indicating that Paul Martin became the
Prime Minister. Similarly, if KB contains a
generalization of the new information, this
information will be redundant.
The relevance of information in a dynamic
stochastic system degenerates gradually over time.
The rate of degeneration of information relevance
with respect to a rational decision maker depends on
the probabilities of change as well as on the relative
utilities (Tawfik and Khan, 2005). Cyc supports a
notion of possibility akin to probability. However, it
is unlikely that the probabilistic knowledge in the
KB will be specified fully to construct dynamic
belief networks. Therefore, we rely on the
intersection of the temporal extents associated with
temporal object in the KB to verify the mutual
relevance of temporal objects. Similarly, most
spatial effects also weaken with distance. Therefore,
it is fair to filter out new knowledge whose spatial or
temporal effects lie outside the scope of interest.
New knowledge could be divided into rules and
events (facts). We consider that the chance seeker
relies on a rule if chance seeker includes some
actions that are causally supported by the
consequences of the rule into its plan. The impact of
the rule measures the role of the rule in reaching the
goals. It could be regarded as the utility changes that
are credited to the rule B. If S represents the state of
chance seeker’s attributes, then impact is given by:
To assess V(S
), we consider two cases: In one
case, V(S
) may already be stated clearly in the rule.
For example, the time saving from taking a newly
built high speed train to a certain destination will be
clearly stated in the news. On the other hand, if
) is unclear, we can deduce a reasonable
hypothesis by combining the new rule and existing
rules in background KB. This hypothesis will not go
beyond the known knowledge. For example, if there
is an assertion in KB stating that all the people in the
same country speak the same language, then
communicating with all Brazilians will be the utility
of learning Portuguese for a chance seeker who
wants to travel to Brazil. Note that this utility could
be inaccurate since it is based on a hypothesis. In
general, impact
may act as a greedy measure of
progress towards the goals but does not guarantee
reaching these goals. An exogenous rule may
undermine actions in the other part of chance seeker.
When new knowledge is an event, to determine
the value of an event, we have to take other factors
into account. An event could be composed by a
bundle of assertions describing its features, such as
actions, locations, time, physical object involved,
etc. The impact of an event according a particular
chance seeker is based on the following features:
Importance of the entities involved in the event.
To evaluate an event, we take the importance of
those objects into account. For example,
‘Microsoft’ may be considered to be a more
important company than other small companies.
However, a small company currently working
with Microsoft may be important.
The relationship between involved objects and
chance seeker needs to be taken into account.
For example, a company owned by family
members may mean a lot to chance seeker
though it’s a small company. For example, the
chance seeker may work for this small business.
Generally, close relatives, friends, and
acquaintances are more important that strangers.
According to the above:
Where V
is a value function that takes into
account the importance/size of objects
the attributes
involved and the relationships between objects and
the chance seeker including spatio-temporal
relationships. V
tries to guess the potential change
in the chance seeker’s attributes.
A negative impact indicates that the new
knowledge is a potential threat. In the case of
irrelevant new knowledge, the impact will be inside
the range of [negative threshold, positive threshold].
The new knowledge will be integrated into KB for
future reference. On the other hand, the new
knowledge will be considered as a chance candidate
if the impact is outside the range.
4.2 The Magnitude of Chances
Here, B is the set of new knowledge that passes
the relevance tests, the system will try to revise
current plans (CP) of the chance seeker using B.
Partial Order Planning (POP) and SATplan
algorithm (Russell and Norvig, 2002) can be used to
generate new plans (NP
) by taking B into account.
In our system, SHOP (Nau et al. 1999) generates
the plans for the chance seeker. SHOP is a domain-
independent automated-planning system. It is based
on ordered task decomposition, which is a type of
Hierarchical Task Network (HTN) planning.
By adopting NP
instead of CP, the chance
seeker may be able to achieve a different set of
goals, or save less time and/or money while
achieving the same goals. All these features can be
reflected by a utility function mapping. The
magnitude of B denoted by M
is represented as the
utility difference between NP
and CP.
There could be a gap between the goals of NP
and the goals of CS. As describing in section 2, a set
of goals can be represented by a future status of
attributes important to the chance seeker. If we use a
utility function (V) to map those attributes into real
values and add them together, we can represent a
notion of preference. The change in the utilities
could be represented as:
represents the difference between new plans
and current plans. If M
in the range of [negative
threshold, positive threshold], it means that NP
CP are roughly the same. The magnitude of B is low.
Whether B is a chance or not, there are the following
possible cases:
Short-term setback: When B has negative effect
on chance seeker’s attribute and no threat to the
current plans, B will be ignored.
Potential risk: When B has negative effect on
chance seeker, and threatens some of the current
plans. However, repair plans can be found such
that the new plans including the repair plans can
achieve the same goal as before. This is
considered a potential risk even though it is
possible to repair the plans because if the
)),(),(( CSObjectrelationsObjectsSizeVimpact
chance seeker proceeds with the original plans
the goals may not be reached.
Risk: Repair plans cannot be found, NP
achieve fewer goals than before. M
is out of
range. The system will consider B is a risk.
Short-term prosperity: When B has positive
effect on chance seeker’s attribute, and no effect
on the current plans.
Exploitable efficiency: NP
can achieve the
same goals as CP but in significantly shorter
time or costs less. B is considered as a chance.
Improved reliability: NP
can achieve the same
goals as before for approximately the same cost
but offer an alternative for some plan elements.
Inefficient alternative: Exploiting B, NP
achieve fewer goals than before or the same
goals at a higher cost without threatning CP. B
is ignored.
Opportunity: NP
can achieve more goals than
before. M
is significant and positive and B is
considered a chance.
Short-term gain long-term risk: When B has
positive effect on chance seeker, threatens some
of the current plans and the plans cannot be
Short-term loss long-term gain: B results in an
immediate loss but enables longer term plans.
Finally, if a chance is detected, NP
will be set
as CP.
4.3 Visualizing Chances
When a chance is detected, visualizing chances is
important as the last step of chance discovery.
Sometimes chance seeker may not understand why
chances returned by chance discovery system are
chances. Visualization of chances could emphasize
on the explanation and help chance seeker to realize
A detail visualization explanation including
display of the future status of attributes of chance
seeker, display of chance seeker’s current plans, etc,
may be necessary. Kundu et al. (2002) present a 3-D
visualization technique for hierarchical task network
plans. Such visualizations will be useful for the
chance seeker to understand the interactions between
various elements in the plan.
The evaluation of chance discovery (CD) systems
could be based on precision, efficiency and chance
management. As discussed in Section 1, many
previous CD approaches regard chances as unknown
hypothesises, focusing on techniques to derive
common chances, i.e. chances for all people. Our
approach focuses on knowledge management,
finding chances in known knowledge (news, WWW,
etc) for a particular chance seeker by the support of
a large and rich knowledge base. In the 2005
tsunami tragedy, scientists correctly detected the
occurrence of the tsunami, but failed to warn the
relevant people in South Asia in time to evacuate.
Hence, chances are relative.
KeyGraph, as introduced in Section 1, is a
widely used technique in CD research. Matsumura
and Ohsawa (2003) present a method to detect
emerging topic (web page as chance) by applying
KeyGraph on web pages. A “Human Genome
project” example was presented. Its benefits include
finding cures to conquer fatal illness. Two sets of
web pages (C
and C
), each containing 500 web
pages, were obtained by searching “human genome”
in Google. C
was obtained on Nov 26, 2000. C
was on Mar 11, 2001. In the output of KeyGraph,
Celera (, a growing HG research
website, was detected as a chance in C
Celera co-occurred with the most important
(foundation) websites in C
. The set of foundation
websites of C
and C
, such as NCBI (the National
Centre for Biotechnology Information), etc, is
almost the same. The following events about Celera
were reported in the meantime:
1. The Human Genome Project team and Celera
announced the completion of the draft sequence
of the human genome in June, 2000.
2. Craig Venter, President and Chief Scientific
Officer of Celera and Francis Collins, Director
of the Human Genome Project, met President
Bill Clinton and British Prime Minister Tony
Blair for the progress of the human genome
3. Papers about the completion were published in
Nature and Science in 2001.
For a researcher in medicine whose goals include
finding a cure for genetic diseases, our CD system
would report a chance after evaluating events 1&2
and would propose new plans. The system may draw
the researcher’s attention to the draft sequence as
early as on Jun 27, 2000 because Clinton and Blair
are very important individuals. The degree of
relevance will be high. The magnitude of “the draft
sequence” will be high since it makes the
researcher’s unattainable goals achievable.
Therefore, our approach could discover chances fast.
This paper describes a chance discovery system
based on Cyc Knowledge base. The knowledge base
works as a virtual reality. Cyc KB simulates the
development of real society by continuously
updating its knowledge. The new knowledge comes
from newspaper, magazine, and WWW, etc. The
chance discovery system searches chances in KB for
on behalf of the virtual chance seekers. By assessing
the relevance of new knowledge, the irrelevant
knowledge to a chance seeker is ignored. Then
chance in relevant knowledge is detected by
considering its impact on the current plans and the
possibility of new plans that are built based on the
new knowledge.
Abe, A., 2003a.. Abduction and analogy in chance
discovery. In Ohsawa, Y. and McBurney, P., editors,
Chance Discovery, pages 231–247. Springer-Verlag
Berlin Heidelberg.
Abe, A., 2003b. The role of abduction in chance
discovery. New Gen. Comput., 21(1):61–71.
Kundu, K., Sessions, C.,,desJardins, M. and Rheingans, P.
2002. Three-dimensional visualization of hierarchical
task network plans, Proceedings of the Third
International NASA Workshop on Planning and
Scheduling for Space, Houston, Texas.
McBurney, P. and Parsons, S., 2003. Agent
communications for chance discovery. In Ohsawa, Y.
and McBurney, P., editors, Chance Discovery, pages
133–146. Springer-Verlag Berlin Heidelberg.
Nau, D. Cao, Y., Lotem, A. and Muñoz-Avila, H. 1999.
SHOP: Simple Hierarchical Ordered Planner. In
Proceedings of the International Joint Conference on
Artificial Intelligence (IJCAI-99), pp. 968-973.
Ohsawa, Y., 2003a. Keygraph: Visualized structure among
event clusters. In Ohsawa, Y. and McBurney, P.,
editors, Chance Discovery, pages 262–275. Springer-
Verlag Berlin Heidelberg.
Ohsawa, Y., 2003b. Modeling the process of chance
discovery. In Ohsawa, Y. and McBurney, P., editors,
Chance Discovery, pages 2–15. Springer-Verlag
Berlin Heidelberg.
Ohsawa, Y., Benson, N. E., and Yachida, M. ,1998.
Keygraph: Automatic indexing by co-occurrence
graph based on building construction metaphor.
Proceedings of the Advances in Digital Libraries
Conference, pages 12–18. IEEE Computer Society.
Ohsawa, Y. and McBurney, P., 2003. Preface. In
Ohsawa,Y. and McBurney, P., editors, Chance
Discovery. Springer-Verlag Berlin Heidelberg.
Ohsawa, Y. and Yachida, M., 1999. Discover risky active
faults by indexing an earthquake sequence.
Proceedings of the Second International Conference
on Discovery Science, pages 208–219. Springer-
Verlag.,2002. OpenCyc documentation.
Parsons, S. and McBurney, P., 2003. Logics of
argumentation for chance discovery. In Ohsawa, Y.
and McBurney, P., editors, Chance Discovery, pages
150–165. Springer-Verlag Berlin Heidelberg.
Russell, S. and Norvig, P., 2002. Artificial Intelligence: A
Modern Approach. Prentice Hall Series in Artificial
Intelligence, 2nd edition.
Tawfik, A. Y. 2004. Inductive reasoning and chance
discovery. Minds and Machines, Volume 14 (Issue
4):441– 451.
Tawfik, A. Y. and Khan, S., 2005. The Degeneration of
Relevance in Dynamic Decision Networks with Sparse
Evidence, Applied Intelligence, to appear.