BIOLOGICAL
CONCEPT FORMATION GRAMMARS
A Flexible, Multiagent Linguistic Tool for Biological Processes
Veronica Dahl
1,2
, Pedro Barahona
3
, Gemma Bel-Enguix
1
and Ludwig Krippahl
3
1
Research Group on Mathematical Linguistics, Rovira i Virgili University, Avda. Catalunya 35, Tarragona Spain
2
Logic and Functional Programming Group, Simon Fraser Unviersity, 8888 University Dr., Burnaby, Canada
3
Centria - Centro de Intelig
ˆ
encia Artificial, Departamento de Inform
´
atica, Universidade Nova de Lisboa, Portugal
Keywords:
Biology, Cognitive sciences, Concept formation, Multi-agent systems, Molecular biology, Nucleic acid string
analysis, Lung cancer detection, Logic programming, Constraint handling rules, Logic grammars, Constraint
handling rule grammars, Language processing, RNA design.
Abstract:
Constraint based models that are useful for processing biological information have been successfully put
forward recently, e.g. for representing multi-disciplinary biological knowledge in view of cancer diagnosis,
and for reconstructing RNA sequences from secondary structure. Here we generalize such results into a
model for biological concept formation which can interact with heterogeneous agents through constraint-based
reasoning. Our model includes linguistic agents, probabilistic agents for mining nucleic acid, and illness
diagnosis agents. Information is selected automatically as a side effect of (the system) searching through
applicable CHR rules, and automatically transformed when a rule triggers; decisions follow from the normal
operation of the rules, and cognitive structure is given by properties that the concepts a given rule is trying to
relate must satisfy. Moreover the user can declare under what circumstances a given property or properties can
be relaxed. Concepts formed under relaxed properties result in output which signals not only what concepts
were formed, but which of the properties associated with the construction of those concepts were satisfied and
which were not. This allows us human-like flexibility while maintaining direct executability.
1 INTRODUCTION
Computational linguistics traditionally breaks up the
tasks of processing language into separate compo-
nents which are often applied in sequence: first a lex-
ical analysis annotates the input with word category
information, then a syntactic parser produces a parse
tree or graph, next a semantic component translates
the parse tree into a meaning representation form, and
so on.
Sometimes some of these components intermin-
gle, e.g. syntactico-semantic parsers build structure in
parallel with meaning representations, in particular to
exploit the fact that semantics can inform syntax, and
viceversa. In such cases, the semantic information is
either embedded into the grammar rules themselves,
or forms part of a separate component, for instance a
taxonomy, which the grammar rules consult.
Even in cases in which the components are sepa-
rate, the master-servant modality of traditional com-
puting science reigns, in that for instance grammar
rules might invoke a taxonomy but never allow the
taxonomy itself to decide when to intervene.
Modern language models and systems, in contrast,
stress interactions between modularized components
of a grammar which collaborate in a more democratic
way. In this respect they approach modern software
agents, which must act in a relationship of agency,
that is, they must perform an action on behalf of an-
other program module, unsolicited. In fact there have
been several recent proposals to explicitly incorporate
agents into grammars, in view of specific applications
such as generating architectonical designs (Grabska
et al., 2009), or designing 3-D structures (Jacob and
Mammen, 2009).
An interesting recent development in logic
grammars- Constraint Handling Rule Grammars, or
CHRG (Christiansen, 2005)- propitiate the introduc-
tion of agents, in that they allow multi-headed rules to
drive the parsing process in a daemon-like fashion: if
all heads in a rule have matching counterparts within
the working store (also called constraint store), then
the rule applies, making no hierarchical distinction
among the different heads of a rule, which contribute
388
Dahl V., Barahona P., Bel-Enguix G. and Krippahl L. (2010).
BIOLOGICAL CONCEPT FORMATION GRAMMARS - A Flexible, Multiagent Linguistic Tool for Biological Processes.
In Proceedings of the 2nd International Conference on Agents and Artificial Intelligence, pages 388-394
DOI: 10.5220/0002786203880394
Copyright
c
SciTePress
all equally. Thus, a CHRG rule can deal for instance
with both syntactic and semantic information in one
stroke, by having syntactic and semantic agents coex-
ist as heads in the same rule.
This capability can be taken even further, since
CHRG rules can pick up working store elements com-
ing from a variety of disparate sources, and thus
they lend themselves ideally for incorporating multi-
agents that collaborate in tasks that require intelligent
interactions with non-grammatical kinds of agents.
In recent years, CHRG models that are useful
for processing biological information have been suc-
cessfully put forward, e.g. for representing multi-
disciplinary biological knowledge in view of cancer
diagnosis (Barranco-Mendoza et al., 2004), and for
reconstructing RNA sequences from secondary struc-
ture (Bavarian and Dahl, 2005).
In this article we generalize such results into a
model for biological concept formation which can in-
teract with heterogeneous agents through constraint-
based reasoning, in view of fine-tuning the process
to obtain richer and more accurate results. These
results include traditional parsing results such as a
sentence
0
s structure and meaning representation, but
also less (for a parser) conventional results such as
mined nucleic acid information, or medical diagno-
sis. In this sense, our model can be viewed as an ex-
ecutable, multi-agent based cross between a grammar
and a biological expert system. It consists basically
of a Constraint Handling Rule grammar that incor-
porates agents for syntax and semantics, but also for
domain dependent biological information, including
probabilities, and allows these agents to interact with
grammatical information in a natural and expressive,
while effective, way.
2 MOTIVATION
Intelligent systems such as those we aim ultimately at
with our new paradigm must exhibit distributed intel-
ligence, since they must represent and consult knowl-
edge from different disciplines (e.g. linguistics, biol-
ogy, parsing). Such knowledge is naturally, and more
manageably, expressed in different modules, one or
several for each discipline. However these modules
must communicate between them in effective ways.
This is no easy task, as anyone who has interacted
with human experts in view of interdisciplinary col-
laboration will attest to. Where computers are in-
volved, it becomes even more challenging.
Even within the one discipline of computational
linguistics, the variety of subareas involved presents
an already formidable challenge. Speech analysis re-
quires different specialized knowledge, for instance,
than text processing.
We typically divide in order to reign, and in many
cases it is straightforward to do so because the pro-
cesses involved can be done independently. For in-
stance, we can first glean text from speech through
one of the many currently available speech processors
out there, and then utilize a text processing program
to mine the text appropriately for our needs. In this
article we will assume speech has already been trans-
lated into text, and focus on text processing.
Even by thus restricting our problem, however, we
cannot always divide to reign in the sense of identi-
fying independent processess that can be run in se-
quence. As mentioned in the Introduction, it may
be desirable to intermingle some of the components.
Thus, syntactico-semantic parsers build structure in
parallel with meaning representations, so that seman-
tics can inform syntax, and viceversa.
As an example of syntax informing semantics,
when a sentence contains several natural language
quantifiers, it is often easier to determine what their
respective scopes -and hence, the sentence
0
s final
meaning- should be once an entire syntactic structure
is initially produced and can be looked at as a whole.
As an example of semantics informing syntax, se-
mantic types associated to lexical items may help dis-
ambiguate what would otherwise be a structurally am-
biguous sentence. For instance “find the price of a
computer having Java” has a sensible reading, namely
“find a computer that has Java and then calculate its
price”, as opposed to the nonsensical (to us, but not
to machines) interpretation “find a computer
0
s price
such that this price has Java”. The nonsensical read-
ing can be disallowed if we only incorporate the se-
mantic type information that computers can have Java
whereas prices cannot. Of course, if we allow the
nonsensical reading it will fail at knowledge-based
consultation time, but disallowing it at the parsing
stage is more efficient than allowing the parser to gen-
erate a “meaning” representation which we know is
doomed to failure.
Semantic type information, however, belongs
epistemologically (and also from a practical point
of view, since it can be consulted per se) in the
knowledge base component of a traditional question-
answering system, and therefore needs to be con-
sulted from its grammar component for the purpose of
blocking non-sensical readings- unless the type infor-
mation were to be redundantly included in the gram-
mar as well, as some systems dealing with semantic
type checking do.
However in our multi-agent model, a single agent
containing all the semantic type information of a par-
BIOLOGICAL CONCEPT FORMATION GRAMMARS - A Flexible, Multiagent Linguistic Tool for Biological Processes
389
ticular application is enough, as it can be consulted
either per se (e.g. for replying to questions such as:
“Is a bank a financial institution?”) or as part of an in-
put sentence
0
s semantic correctness analysis process.
In fact, it can be consulted, if desired, from the same
grammar rule for both purposes.
More importantly, it can pop up in daemon-like
fashion whenever needed, as opposed to depending
on e.g. the parser for consulting it explicitly. For
instance, the lexical entry for “price” could throw
in the workspace not only the grammatical informa-
tion that “price” is a noun with meaning representa-
tion “price(X-T,P)” (the price of an individual X of
type T is P), but also the pragmatic constraint: “con-
sistent(T,price)” (i.e., X
0
s type T must be consistent
with X having a price). Likewise, the lexical entry
for “computer” would attach the appropriate value to
T (namely, “computer”), and once both are in, a bi-
headed rule would provoke a failure if T were not the
type of an object allowed to have a price.
Such issues, which are mostly (computationally)
linguistic in nature, first motivated our quest for a
cognitive science formalism that could abstract the
problems into a general paradigm, concept formation,
that could be directly executable and yet flexible. Our
incursions into computational molecular biology and
other biological applications led, as we next recount,
into interesting extensions of this framework.
3 BACKGROUND: FROM
LINGUISTIC TO BIOLOGICAL
CONCEPT FORMATION
Biological Concept Formation Grammars have
evolved from parsing methods we first developed
specifically for natural language (Dahl and Blache,
2004), then generalized into an executable cognitive
model of knowledge construction inspired as well
in constructivist theory (Dahl and Voll, 2004), and
finally adapted to very different biological process
modeling tasks: early cancer diagnosis (Barranco-
Mendoza et al., 2004), RNA reconstruction from
secondary structure (Bavarian and Dahl, 2005) and
molecular biology text mining (Bel-Enguix et al.,
2009). From each of these applications we have
distilled both common threads and specific idiosyn-
cracies which have served to fine-tune our initial,
general-purpose tool of Concept Formation Rules
into the flexible while specialized biological oriented
model we present in this article.
Succinctly put, Concept Formation exploits the
natural connections (discussed in (Dahl and Voll,
2004)) between constructivism, cognitive logic and
logic programming
0
s recent new paradigm, Constraint
Handling Rules (Fruhwirth, 2002), to develop a cog-
nitive model of knowledge construction which can be
directly executed through (a specialized system im-
plemented in) CHR. In this model, information is se-
lected automatically as a side effect of (the system)
searching through applicable CHR rules, and auto-
matically transformed (or simply augmented) when
a rule triggers; hypotheses can be made in the form
of assumptions; decisions follow from the normal op-
eration of the rules, and cognitive structure is given
by properties that the concepts a given rule is trying
to relate must satisfy. Moreover some latitude is pro-
vided by which rather than rigidly having to satisfy
all properties defined as necessary for a concept to
form, the user can declare under what circumstances
a given property or properties can be relaxed. Con-
cepts formed under relaxed properties result in output
which signals not only what concepts were formed,
but which of the properties associated with that con-
cepts construction were satisfied and which were not.
This allows us human-like flexibility while maintain-
ing direct executability.
3.1 Concept Formation Rules: The
Basic Formalism
Human mind can be seen as a dynamically evolv-
ing store of Knowledge which constatly updates it-
self from new information built from previous infor-
mation with some kind of reasoning (Dahl and Voll,
2004). Concept formation can be defined as the pro-
cess of constructing new pieces of knowledge from
previously known ones, process that might roughly
look as follows: c
1
, c
2
, ..., c
i
newc.
From the perpective of CHR, new pieces of
knowledge are the Body of the rule, and the pre-
viously known ones the Head. Concept Formation
Rules have the same general form as CHR rules, ex-
cept that the guard may include any number of prop-
erty calls for properties which have been defined by
the user:
Head ==> Guard | Body
Head and Body are conjunctions of atoms and
Guard a test constructed from (Prolog) built-in or
system-defined predicates, including the reserved bi-
nary predicate “prop” (for “property”); the variables
in Guard and Body occur also in Head; if the Guard
is the constant true then it is omitted together with the
vertical bar. Its logical meaning is the formula
(Guard (Head Body)
and the meaning of a program is given by conjunction.
ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence
390
Vagueness is expressed by relaxing properties be-
tween concepts in accordance with a user’s criteria.
The criteria can be flexibly and modularly adjusted for
experimental purposes while maintaining direct exe-
cutability.
Thus, rather than inflexibly allowing for a concept
to be formed if a test succeeds and disallowing its for-
mation if that test fails, we single out those tests for
which we want to allow flexibility as properties. Prop-
erties are like any other test, except that their failure
does not result in the rule itself necessarily failing: the
concept will still be formed, and two lists will be asso-
ciated with it: a list of the properties that the concept
satisfies (S) and a list of those which it violates (V).
This allows us to construct possibly incorrect
or incomplete concepts, plus the information on in
which way they are not totally warranted. The user
then has all the information pertaining to the construc-
tion of a particular concept and can therefore interpret
these results in a much more informed, holistic way
than if the degree of randomness or vagueness had
been blindly computed from those assigned a priori
to each individual piece of a reasoning puzzle.
Although the lists of satisfied (S) and unsatisfied
(V) properties are not explicit, they are managed by
the system. The notion of vaguenes lies in the distri-
bution of the lists.
For instance, if we want to accept incorrect sen-
tences in a parsing system that checks for number
agreement, we might designate as properties all tests
on rule applicability that would correspond to correct
parses, and then relax some of them (e.g. the num-
ber agreement property). This would be useful for in-
stance in a second language tutoring system which al-
lows the user to make certain types of mistakes, while
pointing out the reason why those are mistakes (i.e.,
which properties are not being satisfied).
The property calls are automatically handled by
the system provided that the user defines the proper-
ties as follows:
a) a property must be named and defined through
the binary predicate prop, whose first argument is the
property’s name and whose second argument is the
list of arguments involved in checking, and in sig-
nalling the results of checking, the property. For in-
stance, in a grammar that needs to check for number
agreement between a determiner and a noun, say, and
to produce either the (agreeing) number of both, or
an indication of mismatch, we can choose the name
agreement for the property, and define it as follows:
prop(agreement,[Ndet,Nn,N]):- Ndet=Nn,
!, N=Nn.
prop(agreement,[Ndet,Nn,mismatch]).
b) Acceptability of a property that has thus
been defined must be checked in the concerned rule
through the binary system predicate “acceptable”,
whose first argument is the prop atom with all its ar-
guments and whose second argument will evaluate to
either true, false, or a degree of acceptability, accord-
ing to whether (or how much of) the property is satis-
fied. For our example, we can write:
determiner(Ndet), noun(Nn) = =>
acceptable(prop(agreement,[Ndet,Nn,N]),B) |
noun_phrase(N).
c) In order to relax a property named Name (i.e.
to allow the derivation of concepts that require it but
for which it is not satisfied), we simply write the fol-
lowing:
relax(Name).
Degrees of acceptability can be defined through a
binary version of the relaxing primitive, where L is
the prop atom with all its arguments and D is a mea-
sure of acceptability:
relax(L,D).
A list of satisfied and violated properties, together
with the degree of violation if appropriate, will be out-
put for each property defined in a given CF program.
3.2 Biological Concept Formation Rules
Although it was not emphasized in the first stages of
our work, Concept Formation Rules involve agents,
in the sense of cooperating processes that trigger au-
tonomously upon need, i.e. when the constraint store
acquires (either through a user
0
s query or through the
normal working of the rules) enough data to trigger a
given rule, it will trigger on its own and produce, if all
involved properties hold with an acceptable tolerance
level, the appropriate new concepts to be added to the
working store. The process continues until no more
new concepts can be added.
Since Concept Formation is a Cognitive Sciences
model, note that although our examples so far have
been grammatical in nature, it can also be used to
model any other problem domain.
Having the main mechanism for Concept Forma-
tion, we need to introduce the agents that can help
to design efficienty applications for BioMedicine. In
particular, we can identify three main types of agents:
A Property Agent: which manages the user-
defined properties.
A Concept Formation Agent: which invokes it in
order to enforce or relax those properties accord-
ing to the user
0
s specifications as explained above.
BIOLOGICAL CONCEPT FORMATION GRAMMARS - A Flexible, Multiagent Linguistic Tool for Biological Processes
391
A Probabilistic Agent: which is needed for the
application of formalism to fuzzy domains, like a
multidisciplinary approach to cancer diagnosis, as
well as from our work on RNA sequence discov-
ery from secondary structure.
A BioMining Agent: which can identify substrings
of interest within given strings of nucleic acid.
The definition of such agent derives from our
work on mining molecular biology texts.
In the next section we describe the resulting new
model of concept formation in its rightfully earned
conceptualization as a multiagent system for biologi-
cal concept formation.
4 OUR MODEL
0
S LINGUISTIC
AGENTS
Our model comprises two types of linguistic agent
systems: those that can process human language input
and those which can analyse nucleic acid sequences.
We next explain the general idea of Human Language
processing Agent Systems and describe Nucleic Acid
Language Agent Systems, stressing their probabilistic
agents and suggesting an application to illnes diago-
nis.
4.1 Human Language Processing Agent
Systems
A linguistic agent that can process human language
input allows non-computer specialists to pose ques-
tions directly in natural language, thus largely remov-
ing the need for them to either learn a specialized
computing language, or depend on computer special-
ists who may not be too conversant with the biolog-
ical side of things. The subset of language admitted
by this agent will vary according to the application,
but a core, extensible parser conforms a basic lin-
guistic agent which, because expressed in Constraint
Handling Rule format, allows for smooth interaction
between the grammar and the biological knowledge
base agents (by allowing one of them to inform the
other one through integrity constraints in CHR). The
main features of this agent, described in (Dahl, 2009),
are that it allows for eager discarding of wrong lines
of reasoning, and for paraphrases of a given ques-
tion without ill-effects in the execution, thus accept-
ing fairly rich input.
4.2 Nucleic Acid Language Agent
Systems
In (Bel-Enguix et al., 2009), we proposed a CHR
based mining technique- the Parallel Matching
methodology- for problems that are interesting both in
molecular biology and in linguistics, such as identify-
ing subsequences (of English words, for instance, or
of nucleotides) that are common to a group several se-
quences, matching ambiguous subsequences, finding
a substring
0
s frequency, finding gapped patterns. The
resulting set of primitives can be viewed as conform-
ing a nucleic acid decoding agent, which has more-
over been expanded to include further nucleic acid
decoding primitives, and into a grammatical formu-
lation (Dahl, 2009) allowing interaction with human
language processing agents, so that the nucleic acid
agents can be consulted from human language com-
mands.
Our CHR plus CHRG formulation more readily
allows us to use eagerly any constraints of the prob-
lem which could serve to early prune the search space.
For instance, if the presence of phenylalanine pre-
cluded that of leucine and we had detected phenylala-
nine in our input string, there would be no point in
searching for leucine or for the sequence that encodes
it. An integrity constraint that provokes failure if both
are found in the working store (a single line of code)
is all that is needed. In a straight Prolog formulation,
in contrast, adding something globally would not be
possible: we would have to enter the definition of the
substring finding predicate to include a test in some
appropriate place within it.
4.2.1 Probabilistic Agents
To mine more complex biological structures, such as
RNA secondary structure, in order for instance to re-
construct the RNA sequence that folds into a given
secondary structure, we also use CHR but add a prob-
abilistic agent that follows the methodology devel-
oped by (Bavarian and Dahl, 2005). This agent op-
erates in the guard of a CHR rule, and uses the prob-
abilities that are believed to govern the proportion of
base pairs within RNA sequences. Bavarian and Dahl
calculated these probabilities by comparing several
RNAs together from Gutell lab
0
s comparative RNA
website (Cannone et al., 2002), a database of known
RNA secondary structures. After comparing 100 test
cases with various length from 100 to 1500 bases,
they found the following probabilities for each base
pair:
P
CG
= 0.53, P
AU
= 0.35, P
GU
= 0.12
The other probabilities which are of interest are
the probabilities for an unpaired base to be one of A,
ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence
392
C, G, or U. The results are as follows:
P
G
= 0.18, P
A
= 0.34, P
C
= 0.27, P
U
= 0.21
Inserting the probabilities into the grammar rules
is done by generating a random variable in the guard
section of the rules, which is the only part that accepts
Prolog predicates. This random variable then is tested
according to the probabilities: for instance if the ran-
dom variable in the guard of a rule that assigns nu-
cleotides to positions known to be paired is less than
0.53, it will assign a GC pair. The average error is es-
timated to be about 18%, meaning that 18% of the nu-
cleotides might be paired with a nucleotide in a wrong
position (in the original structure they might be either
unpaired or be paired with another nucleotide).
4.2.2 Illness Diagnosis Agents
In previous work of one of the authors with Alma
Barranco-Mendoza, specialized concept formation
rules were used for representing knowledge in view
of diagnosing diseases such as lung cancer (Barranco-
Mendoza et al., 2004). Following this work, yet an-
other kind of probabilistic agent materializes as an ad-
ditional parameter of each constraint in the special-
ized concept formation rules of our multi-agent sys-
tem.The application introduced in this paper aims to
aid in early stage detection of some types of cancer,
like lung and oral, which have poor prognosis because
they are very difficult to diagnose at the early stages.
Our concept formation methodology assists in the
integration and analysis of multidisciplinary agents
containing genetic and molecular information along
with the radiological, serum and sputum data. In par-
ticular, it provides some kind of diagnosis even if
given incomplete patient information, as not all tests
can or will be done on a given patient at a given time.
This is achieved by relaxing certain properties, where-
upon the analysis will be completed even if the infor-
mation is not complete. The list of violated properties
can provide a list of suggested follow-up tests to im-
prove the accuracy of the diagnosis.
As part of the input concepts it accepts the pa-
tients age, smoking history, malignancy history, ra-
diological, serum and sputum data. The knowledge
store includes the properties that should be evaluated
for each input data element as well as the relations
amongst them. The diagnosis is given as a probability
of cancer that is calculated as a function of the con-
cepts used in the analysis. As well, the diagnosis will
list those diagnostic properties that were satisfied and
those that were not. For example:
const(Prob),age(x,A),history(x,smoker,T),
serum_data(x,marker_type,in_range)<=>
marker(x,marker_type,in_range,P,B),
acceptable(marker(x,marker_type,in_range, P),B),
probability(P,Prob,x, B),
acceptable(probability(P,Prob,x),B)|
possible_lung_cancer(yes,Prob,x).
relax(marker(x,marker_type,in_range,P,B)).
This rule evaluates for a patient x if a specific
biomarker, marker-type, found in serum data is within
a certain value range for a patient with an age of A
who is a type T smoker (T depends on the number of
cigarettes or cigars smoked daily). If true, then the
diagnosis of possible lung cancer is going to be true
with a probability increase of P (where P is a func-
tion of the patient’s age, health history, and this par-
ticular biomarker presence). But if we relax the re-
quirement of the presence of the biomarker, then the
system can evaluate patient records that do not have
this particular information and report in the diagnosis
listing that this information was not included in the
record, which could be valuable information as rec-
ommended follow-up tests for that particular patient.
5 CONCLUSIONS
We have abstracted, from recent different realiza-
tions of the linguistically inspired Concept Formation
paradigm, a multi-agent model for Biological Con-
cept Formation which can be considered as a com-
putational metaphor for the (biological) mind, with
direct executability implications. Due to the general-
ized use of Constraint Handling Rules or their gram-
matical counterpart, we are able to integrate human
language processing techniques into our approach
which are not only useful for all types of concept for-
mation but also allow us a smooth integration of hu-
man language processing agents, as well as their in-
teractions with the knowledge base agents. Another
interesting feature of our proposal is its robustness:
due to the capability of relaxing some of the prop-
erties involved in concept formation, results that can
be useful are provided even in the absence of all the
information “necessary” to form the concepts in ques-
tion.
Concept formation rules are applicable to many
other AI and cognitive problems as well, most no-
tably, those involving the need to reason with incom-
plete or incorrect concepts.
The flexibility allowed by relaxing properties was
argued in our initial paper (Dahl and Voll, 2004) to
provide a more appealing solution to the need for
flexibility than the two main alternatives out there,
namely probabilities and fuzzy logic. The probabilis-
tic approach had been discounted as inappropriate for
measuring the meaning of information, although ad-
BIOLOGICAL CONCEPT FORMATION GRAMMARS - A Flexible, Multiagent Linguistic Tool for Biological Processes
393
equate to measure the randomness of information.
However, after our work on reconstructing RNA se-
quences from the structure into which they fold, plus
our work on using Concept Formation as an aid for
early diagnosis of lung cancer, we must rectify that
statement. We are now convinced that probabilistic
agents, both of the randomness measure kind and in
the form of combining the probabilities of individual
biomarkers into an overall probability of a disease de-
veloping, are very useful agents that have a rightful
place in a biological rendition of the Concept Forma-
tion paradigm.
Admittedly, the range of disparate biological
problems we have attempted to cover under a sin-
gle paradigm is perhaps too ambitious to allow us
as homogeneous a model as we would have liked.
However we feel it is an important step towards
achieving an encompassing and robust multi-agent
model for these various tasks, in that it allows for au-
tonomous triggering of the agents needed at a given
time, for easy synchronization between the various
agents, mainly through integrity constraints, and for
flexibility, through property relaxation, in the face of
either incomplete or erroneous data- a problem that
biological systems aspiring to deal with real life prob-
lems must absolutely face.
ACKNOWLEDGEMENTS
This paper is supported by the European Commis-
sion in the form of V. Dahl
0
s Marie Curie Chair
of Excellence, the Canadian National Sciences Re-
search Council, the project “Constraint - and Hypo-
thetical - based reasoning for BioInformatics”, refer-
ence HP2008-0029, and “Logic, Automata and For-
mal Languages”, MTM2007-63422.
REFERENCES
Barranco-Mendoza, A., Persaoud, D., and Dahl, V. (2004).
A property-based model for lung cancer diagnosis. In
RECOMB (2004) 27–31.
Bavarian, M. and Dahl, V. (2005). Rna secondary structure
design using constraint handling rules. In WCB’05,
Workshop on Constraints for Bioinformatics.
Bel-Enguix, G., Jim
´
enez-L
´
opez, D., and Dahl, V. (2009).
Dna and natural languages: Text mining. In Proc.
International Joint Conference on Knowledge Discov-
ery, Knowledge Engineering and Knowledge Manage-
ment, KDIR 2009, pages 140–145.
Cannone, J., Subramanian, S., Schnare, M., J.R. Collett,
L. D., Du, Y., Feng, B., Lin, B., Madabusi, L., Muller,
K., Pande, N., Shang, Z., Yu, N., and Gutell, R.
(2002). The comparative rna web (crw) site: An on-
line database of comparative sequence and structure
information for ribosomal, intron, and other rnas. In
BioMed Central Bioinfo. 3:2.
Christiansen, H. (2005). CHR grammars. In Journal on
Theory and Practice of Logic Programming, vol. 5,
number 4-5.
Dahl, V. (2009). Decoding nucleic acid strings through hu-
man language. In Manuscript Submitted for Publica-
tion.
Dahl, V. and Blache, P. (2004). Directly executable con-
straint based grammars. In Proc. Journees Fran-
cophones de Programmation en Logique avec Con-
traintes.
Dahl, V. and Voll, K. (2004). Concept formation rules: An
executable cognitive model of knowledge construc-
tion. In NLUCS’04, International Workshop on Natu-
ral Language Understanding and Cognitive Sciences.
Fruhwirth, T. (2002). Theory and practice of constraint han-
dling rules. In Special Issue on Constraint Logic Pro-
gramming (P. Stuckey and K. Marriot, Eds.), Journal
of Logic Programming.
Grabska, E., Strug, B., and Slusarczyk, G. (2009). A multi-
agent distributed design system. In 7th International
Conference on PAAMS’09, AISC 55. Springer-Verlag
Berlin Heidelberg.
Jacob, C. and Mammen, S. V. (2009). Swarm grammars:
growing dynamic structures in 3d agent spaces. In
Digital Creativity, Volume 18, Issue 1, March 2007.
Routledge.
ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence
394