Theodore Andronikos.

International Journal of Engineering Research and Applications

www.ijera.com
ISSN: 2248-9622, Vol. 10, Issue 11, (Series-I) November 2020, pp. 23-30

RESEARCH ARTICLE OPEN ACCESS

Complex Matrices for the Approximate Evaluation of

Probabilistic Queries
Theodore Andronikos*
*(Department of Informatics, Ionian University, 7 Tsirigoti Square, Corfu, Greece

ABSTRACT
This paper studies the evaluation of probabilistic SPARQL queries. The evaluation of such queries is, in
principle, conceptually simple and straightforward. However, it does incur a time cost that must be taken into
account. It thus expedient to devise methods that lower this cost. We propose and explain such a method, which
makes use of complex matrices. The idea to resort to the complex numbers is inspired from the greatly
expanding field of unconventional computing, and, in particular, from quantum computing. This novel proposal
simplifies and speeds up calculation of products of probabilities. Therefore, it is particularly promising in these
cases where the same answer set can be obtained either by employing exact computations or by employing
suitable approximations, as is indeed the case in many probabilistic queries.
Keywords - RDF graph, SPARQL query, probabilistic SPARQL query, complex matrices
----------------------------------------------------------------------------------------------------------------------------- ----------
Date of Submission: 28-10-2020 Date of Acceptance: 09-11-2020
----------------------------------------------------------------------------------------------------------------------------- ----------
of transitive predicates, makes possible the
I. INTRODUCTION formation of paths resembling those typically found
During the last two decades numerous in standard directed graphs. The paths start from an
researches have focused their attention to all things initial node, follow adjacent directed edges labeled
related to the Web. This tremendous effort has by R, and finally terminate at some other node. It is
produced state of the art technologies and has implicitly assumed that such a path expresses some
literally embedded the World Wide Web into meaningful property and that this property can be
everyone’s life. The resounding success of the whole formulated by SPARQL.
endeavor can be, at least in part, attributed to the Let us now shift the emphasis to the fact
adherence to well-designed standards. Linked Open that in many situations that data contained in the
Data [1] provides guidelines regarding the storage dataset are approximations, estimations, beliefs, i.e.,
and communication of data on the Web. The they lack certainty. The world is uncertain and, in
Resource Description Framework (RDF) is the most many respects, probabilistic. Therefore, the data
prominent standard governing the storage of stored in an RDF graph may not be totally correct,
information, whereas SPARQL deals with retrieving but probably correct. In domains where uncertainty
this information, i.e., querying the data. is prevalent, we may adopt a new perspective,
RDF promotes the use of directed graphs namely that the triplets are assigned a nonnegative
[2] as convenient and effective structures for real value that indicates the probability of being
keeping information. The data contained in the RDF accurate. It is expedient to assume that this number
graph obey an explicit syntactic pattern: subject  takes values in the interval [0, 1], as expected from a
predicate  object. The idea behind this is simple probability measure. Alternatively, one may well
and functional: the predicate relates the subject with view this number as a weight or even as a degree of
the object. The rules by which one can query such certainty. There are many domains of practical
datasets are stipulated by SPARQL [3]. This work is importance where probability and uncertainty arise.
concerned with probabilistic Regular Path Queries, One important example worth mentioning would be
which have the potential to specify a path of biological data where connections among biological
adjacent nodes in the underlying graph. This is concepts have a probabilistic nature and where
achieved through transitive predicates. To see how additionally the links themselves are statistically
this works, let us start by picturing predicates as independent [4]. It is, thus, evident that tools and
labels on the directed edges of the RDF graph. In techniques must be developed to address the issue of
this setting, a predicate R is defined to be transitive uncertainty.
if two triples (x, R, y) and (y, R, z) lead to the
inference of the triple (x, R, z). Hence, the existence

www.ijera.com DOI: 10.9790/9622-1011012330 23 | P a g e

Theodore Andronikos. International Journal of Engineering Research and Applications
www.ijera.com
ISSN: 2248-9622, Vol. 10, Issue 11, (Series-I) November 2020, pp. 23-30

They also examined certain algorithmic issues and

1.1 Related work proposed two pruning algorithms. Schoenfisch in
In this subsection, we shall refer to some [17] examined probabilistic ontologies, as well as
notable works that are relevant to this paper. This the computation of queries on such ontologies.
connection should be understood either in the sense Krompass et al. in [19] studied more general
that they study theoretical aspects of the RDF and probabilistic databases. For further insight the
SPARQL formalisms, or that they propose interested reader is also referred to [19] and
mechanisms that enhance the capabilities of RDF references therein.
and/or SPARQL. Zhang et al. in [6] studied path The query language pSPARQL was
queries making use of context-free grammars, which proposed in [20] by Fang and Zhang, and was
are strictly more powerful than regular expressions. further elaborated by Fang in [21]. This language is
Their primary motivation was the restricted an important effort towards the ultimate goal of
expressiveness of the latter, for example in the case probabilistic query evaluation. It subsumes vanilla
of same generation-queries. They proposed an SPARQL and contains additional constructs to
enhanced query language, aptly named cfSPARQL, handle probabilistic queries. An approach that
which is SPARQL with context-free capability. focuses on the evaluation of probabilistic queries
cfSPARQL, being more expressive than vanilla through the use of probabilistic automata is outlined
SPARQL, allows for more interesting queries. Sistla in [22] and in [23]. The current paper is an extension
et al. in [7] showed that queries can be understood in of these works, but with two major differences. First,
terms of automata, and, as a matter of fact, can be there is a shift in the perspective from exact
formulated using techniques from automata theory. evaluation to approximate evaluation. We argue that
A similar approach was followed in [8] and [9], this is acceptable in a situation involving products of
where Büchi automata and ω-automata were probabilities, as will be seen in the following
associated to queries on Linked Data and Path sections of this paper. Second, and, perhaps, more
Queries, respectively. In the same way, Wang et al. importantly, we do not use real numbers, but turn to
in [10] developed a methodology to answer queries complex numbers. This is inspired from another
based on automata. recent prevalent trend, that of unconventional
The algebraic characterization of SPARQL computing. One aspect of unconventional computing
queries was undertaken by Schmidt et al. in [11]. is quantum-inspired computation. Mimicking
The classified queries into equivalence classes and techniques and concepts from quantum computing
established connections with complexity theory. A can provide new insights into classical problems (see
different attempt at query classification was [27], [28] and [29] for an application in optimization
conducted in [12], where another, quite distinct, problems). Often, replacing a matrix of real numbers
notion of equivalence between SPARQL queries was by a matrix of complex numbers satisfying some
defined and studied. The possibility of adding extra property can give better results, as verified
additional information (of a temporal nature) was experimentally (see for instance [24], [25] and [26]
also discussed in [14]. for the use of unitary matrices instead of typical real
The importance of incorporating matrices, as well as extensive comparative test
probabilities in the RDF datasets and evaluating results).
probabilistic queries has, by now, gained widespread
acceptance in the research community. An 1.2 Contribution
influential position paper by Reynolds [16] stressed The contribution of this paper lies on its
the practical usefulness of dealing with uncertain novelty. It demonstrates convincingly that
data and gave pointers about possible solutions that probabilities can be approached via complex
can be employed to overcome difficulties regarding numbers, and in particular numbers of the form ,
their implementation. Huang and Liu in [4] where , which of course immediately implies
developed an approach, which, by suitable extending
that . This facilitates the calculation of
SPARQL, would allow the evaluation of queries on
probabilistic databases. They also presented an products of probabilities, but the end result has to be
approximation algorithm that is able to evaluate carefully interpreted. The result would not be exact,
efficiently path expressions. Likewise, the potential only approximate. Still its usefulness cannot be
to query data characterized by uncertainty has been overlooked, because in a setting of uncertainty
studied by many researchers. Hua and Pei in [13] probabilities close to 0 mean that the assertion is
examined probabilistic path queries using a method virtually improbable and can be dismissed. Hence,
inspired from dynamic programming. Lian et al. in by fixing a threshold, something that is standard
[15] incorporated probability values to typical practice in probabilistic analysis, we can use the
datasets, converting them into probabilistic datasets.

www.ijera.com DOI: 10.9790/9622-1011012330 24 | P a g e

Theodore Andronikos. International Journal of Engineering Research and Applications
www.ijera.com
ISSN: 2248-9622, Vol. 10, Issue 11, (Series-I) November 2020, pp. 23-30

approximate results for valid, with respect to the

agreed threshold, assertions.

II. MOTIVATION
In this section, we shall present an
extensive example to motivate the reader and The matrix represents all the
explain the idea behind this work. information contained in the RDF graph about
Example 1. Suppose we are given the RDF predicate R and is a representation of obj1. The
graph shown in Figure 1a. The predicate R is a element of stores the probability value for
transitive predicate and the numbers are
the triplet . A value of zero indicates
probabilities.
the absence of such a triplet. In this matrix - vector
representation we can compute the products ,
and , which will give us the following
results.

A product of the form contains the

pairs of objects that related via an R-path of length
. The first product provides no more
information than that already shown in the RDF
graph. Things get interesting when we look at
Figure 1a. The RDF graph above shows how and . These vectors contain the probabilities
objects are related through the transitive predicate that obj1 is related to other objects through R-paths
R and the degree of certainty associated to each of length 2 and 3, respectively. As before, the value
such relation. zero signifies the absence of such a path, i.e., that
there is no connection though predicate R. The
SELECT ?x actual numbers, reveal the existence of two inferred
WHERE { triplets and ,
Obj1 R+ ?x . respectively. The former has probability , which
}
makes the corresponding assertion very likely, while
Figure 1b. The above SPARQL query returns the the latter has probability , which makes the
objects that are related to obj1 through an R-path. corresponding assertion ambivalent. So, by taking all
the above results into account, we can be fairly
The triple is assigned certain that the inferred assertion is
probability , which means that it is almost certain true, but cautious about the inferred assertion
that the relation R holds between obj1 and obj4. On . To develop better intuition about
the other hand, the triple is the situation, we can repeat the calculations for
associated with probability , which indicates that vector corresponding to obj2, where
relation R may or may not hold between obj2 and
obj4 with equal probabilities. The query in Figure 1b
uses the syntax of SPARQL 1.1 that makes it
possible to express path properties. The meaning of
the symbol special symbol + is that there exists an
R-path (that is one or more edges labeled by R)
between obj1 and every listed object. Now we observe that the degree of
Let us now see how this query can be certainty regarding the inferred assertion
computed using matrices and (column) vectors. drops to , i.e., we can be rather
Consider the following matrix and vector . sure that is does not hold. These situations always
come up in probabilistic scenarios. In order to
address them practically and efficiently, it is

www.ijera.com DOI: 10.9790/9622-1011012330 25 | P a g e

Theodore Andronikos. International Journal of Engineering Research and Applications
www.ijera.com
ISSN: 2248-9622, Vol. 10, Issue 11, (Series-I) November 2020, pp. 23-30

common practice to fix a cut-off threshold, or cut- and respectively. Let

point (see also [30] and [31]). Then, every assertion us now compute the matrices and .
or fact with probability value below the cut-point is
assumed to be false, and may be ignored in any
further computation. If we assume for a moment that
the cut-point is , then for normalization purposes
all inferred assertion with probability value
should be dismissed, i.e., obj1 is not related to obj5, The matrix represents the triple
and, similarly, obj2 is related neither to obj3 nor to associated with the complex number
obj5. The corresponding vector components can be
. Two things must be emphasized in this case.
taken to be zero. However, increased accuracy may
dictate a smaller threshold, e.g., . Then only those First, computing this number involves just one
addition instead of one multiplication required by
assertions with probability are neglected. This
the computation of . The advantage of using
means that and are
additions over multiplications is very important, as,
now kept and used for the rest of the computation, in a practical setting, this will scale in proportion to
and only the inferred assertion is the size of the RDF graph. However, care is needed
dismissed. in the interpretation of the result. A small error arises
Up to this point, the whole approach was in the result of adding the angles instead. More
conventional. Let us now turn to the unconventional precisely, the probability implied by is
and introduce the complex matrix . , instead of the correct
value 0.81. This can be handled in two ways.
One is to fix a cut-point. In the
conventional setting the cut-point is a small positive
number, e.g., , , even , whose interpretation
is that a probability below that threshold means that
. the inferred fact probably does not hold, and, hence,
should be disregarded. In our new complex setting
This matrix can be constructed from the the cut-point is the exponent of . Specifically, we
RDF graph of Figure 1a, as follows. First, we recall
may fix a cut-point (in radians) such that
the trivial fact that probabilities are real numbers in
the interval . As such, a probability value can whenever we encounter a complex number with
be considered as the of an angle , that is , we understand that it represents a very small
, which allows us to compute the angle likelihood and the associated triple can be ignored.
For example, by setting equal to , or or
as . Having retrieved the angle , we
even we can simulate the cut-points , ,
may go one step further and enter the complex realm
and , respectively. Choosing a cut-point that is
by using the formula . The
suitable for a particular dataset then becomes a
resulting matrix is equivalent to , only now
matter of deciding which is the acceptable degree of
the probabilities have become angles (in radians). In precision for that dataset. If we assume that is
this way, probabilities and correspond to
chosen as an appropriate cut-point, then the
and , respectively. We may now corresponding . Therefore, for the case at
calculate the products and . hand, the particular is less than the cut-
point, which implies that the triple
is considered a valid inference and is accepted.
The other way is to correct the numerical results
using a very simple formula containing a correction
coefficient and the information provided by the sine
of the angle. We propose the formula
As expected and are just the first
, where is the angle in question and
and second column of and contain those objects
is the correction coefficient, as a suitable
explicitly related to obj1 and obj2, along with the
approximation. The philosophy behind the
complex numbers and . The probabilities uncertainty setting is to cope with inexact or
that the triplets and approximate information. In such a situation,
are valid are given by approximating the degree of uncertainty is the only

www.ijera.com DOI: 10.9790/9622-1011012330 26 | P a g e

Theodore Andronikos. International Journal of Engineering Research and Applications
www.ijera.com
ISSN: 2248-9622, Vol. 10, Issue 11, (Series-I) November 2020, pp. 23-30

reasonable approach. Of course, the correction below the adopted cut-point. Then, even when the
coefficient must be chosen carefully. Due to the non- matrix coefficient is the nonzero complex
linearity of the trigonometric functions, it is number (expressing the initial degree of
impossible to use one correction coefficient for the confidence about the corresponding triple), if it
entire interval . The correction coefficient
happens that , then must be replaced by
suitable for a specific subinterval must be validated
.
experimentally. Table 1 below shows that to get a
good approximation for the probability subinterval In the process of answering queries, starting
we may use the formula from the known given matrix , it is inevitable that
with . This formula gives for the some matrix power will be computed. The
approximate value 0.77, which is very close to the coefficients of have the obvious
accurate value 0.81. ▲ interpretation. If , then with approximate
probability there exists an inferred triple
Table 1. The coefficient for the interval [0.8, 0.9]. through an R-path of length or,
equivalently, with approximate probability
there exists an R-path of length from
node to node . If , then there exists no
inferred triple through an R-path of length
or, equivalently, there exists no R-path of length
from node to node . Remark 1 also applies in this
III. MAIN RESULTS case, i.e., if it happens that , where is the
In this section, we shall present the ideas cut-point, then must be replaced by ,
outlined in Example 1 in a general and formal way. something that will add one more zero to the matrix
Definition 1. Let G be an RDF graph . Alternatively, it is worth considering the
fragment containing a transitive predicate R, and let scenario where , which means that
, be an arbitrary enumeration of its nodes.
represents a probability too large to be ignored. In
We associate to G the matrix whose such a case, it is expedient to address the relative
elements are defined as follows: error. We caution the reader here that there is no
error in the coefficients of , since they express
explicit facts taken from the RDF graph itself. The
errors occur when we compute the powers , where
Definition 2. The cut-point is a positive . As we have explained in detail in Example 1,
real number, understood to express radians, such that the following very simple formula (1) can be used
any complex number of the form with , is for this purpose.
simply replaced by (zero).
Remark 1. The choice of the cut-point is of It is more accurate to approximate the
great importance because it has the potential to probability that corresponds to the triple by
significantly facilitate computations by providing a formula (1). The complex number
strict criterion for consistently omitting very small contains the information
probabilities. This policy has as an immediate and regarding , so it only remains to estimate the
concrete practical impact on the measurable speed- correction coefficient . The non-linearity of the
up of the evaluation of probabilistic SPARQL and functions preclude the possibility of
queries. Obviously, datasets of different nature, or
increased accuracy requirements, must be taken into one correction coefficient for all cases. Instead, the
account when determining the cut-point. An probability interval must be partitioned into
aggressive approach for increased performance smaller subintervals. This can be achieved through
could be to set the cut-point , extensive experimental tests. Tables 2a - 2d give the
suggested correction coefficients for the subintervals
corresponding to probability . Such an approach
, , , and ,
might also modify the original complex matrix .
respectively. The Tables also contain indicative
This is due to the fact that in the above Definition 1, examples of probability values within the respective
the case that the triple does not exist includes the
interval, along with the value of before the
case where the probability attributed to a triple is
correction of formula (1) is applied.

www.ijera.com DOI: 10.9790/9622-1011012330 27 | P a g e

Theodore Andronikos. International Journal of Engineering Research and Applications
www.ijera.com
ISSN: 2248-9622, Vol. 10, Issue 11, (Series-I) November 2020, pp. 23-30

probability, that is facts that are highly unlikely to

hold. Another way to deal with this, especially suited
Table 2a. The coefficient for the interval [0.6, 0.7]. to cases where increased accuracy is required, is the
application of the correction formula (1). This
formula has been validated experimentally through
considerable testing, and some characteristics
examples are shown in Table1 and Tables 2a - 2d.

IV. CONCLUSION
This paper advocates the evaluation of
Table 2b. The coefficient for the interval [0.5, 0.6]. probabilistic SPARQL queries via unconventional
means, namely the use of complex matrices. The
extensive example we have presented in section 2
describes in detail the proposed methodology. This
novel approach has an unquestionable advantage,
and that is that the computation proceeds via
additions instead of multiplications. Thus, the
potential to scale down the computational cost is real
and pragmatic. Finally, we have suggested two ways
Table 2c. The coefficient for the interval [0.4, 0.5]. to handle efficiently the numerical errors that will
come-up, either by utilizing a cut-point or my using
an approximation formula, which has been
thoroughly tested experimentally.

REFERENCES
[1] LOD Project, 2014. Linking Open (LOD)
Data Project,
http://esw.w3.org/topic/SweoIG/TaskForces/
CommunityProjects/LinkingOpenData.
Table 2d. The coefficient for the interval [0.2, 0.4]. [2] Resource Description Framework (RDF),
https://www.w3.org/TR/2015/NOTE-rdfa-
primer-20150317/.
[3] SPARQL 1.1 Query Language. Tech. rep.,
W3C (2013),
https://www.w3.org/TR/2013/REC-sparql11-
query-20130321/.
[4] Huang, H., Liu, C.: Query evaluation on
probabilistic RDF databases. In: International
Conference on Web Information Systems
Engineering, pp. 307–320, Springer (2009).
[5] Hartig, O.: An overview on execution
To biggest benefit of the proposed approach strategies for Linked Data queries,
is the fact that the computation of the matrix powers Datenbank-Spektrum 13(2), 89–99 (2013).
involves only the operation of addition (of [6] Zhang, X., Feng, Z., Wang, X., Rao, G., Wu,
angles) and foregoes the operation of multiplication W.: Context-free path queries on RDF graphs.
of real numbers that the conventional approach In: International Semantic Web Conference,
would entail. Using additions instead of pp. 632–648, Springer (2016).
multiplications is always preferable, as the [7] Sistla, A.P., Hu, T., Chowdhry, V.: Similarity
compound computational cost scales down based retrieval from sequence databases using
significantly with the size of the problem at hand. automata as queries. In: Proceedings of the
Hence, this method is undoubtedly better than the eleventh international conference on
conventional. Of course, there is a trade-off, which Information and knowledge management, pp.
in this case takes the form of arithmetic errors. There 237–244, ACM (2002).
are two ways to remedy this situation. One is the [8] Giannakis K., Andronikos T., “Querying
systematic use of a cut-point, a well-established Linked Data and Büchi Automata”, IEEE
technique in probabilistic scenarios, that can safely Proceedings of the 9th International
and rapidly dismiss inferred fact of small Workshop on Semantic and Social Media

www.ijera.com DOI: 10.9790/9622-1011012330 28 | P a g e

Theodore Andronikos. International Journal of Engineering Research and Applications
www.ijera.com
ISSN: 2248-9622, Vol. 10, Issue 11, (Series-I) November 2020, pp. 23-30

Adaptation and Personalization (SMAP), 6-7 [20] H. Fang and X. Zhang, “pSPARQL: a
November, Corfu, Greece, pp. 110 - 114, querying language for probabilistic RDF
2014. (extended abstract),” in Proceedings of the
[9] Giannakis K., Theocharopoulou G., Papalitsas ISWC’16, Posters, 2016.
C., Andronikos T., Vlamos P., “Associating [21] Fang, H. pSPARQL: A Querying Language
ω-automata to Path Queries on Webs of for Probabilistic RDF Data Complexity,
Linked Data”, Engineering Applications of Hindawi Limited, 2019, 1-7.
Artificial Intelligence, Elsevier, Volume 51, [22] Andronikos T., Singh A., Giannakis K.,
May 2016, pages 115-123. Sioutas S., “Computing probabilistic queries
[10] Wang, X., Ling, J., Wang, J., Wang, K., Feng, in the presence of uncertainty via probabilistic
Z.: Answering provenance-aware regular path automata”, Algorithmic Aspects of Cloud
queries on RDF graphs using an automata- Computing, Third International Workshop,
based algorithm. In: Proceedings of the 23rd ALGOCLOUD 2017, Vienna, Austria, 5
International Conference on World Wide September, 2017, Revised Selected Papers.
Web, pp. 395–396, ACM (2014). Springer Theoretical Computer Science and
[11] Schmidt M., Meier M., Lausen G.: General Issues, Volume 10739, pp. 106-122,
Foundations of SPARQL Query ISBN: 978-3-319-74874-0 (Print) 978-3-319-
Optimization. In: Proceedings of the 13th 74875-7 (Online), 2018.
International Conference on Database Theory [23] Andronikos T., Singh A., Giannakis K.,
(ICDT '10), pp. 4–33, Lausanne, Switzerland, Sioutas S., “Computing probabilistic queries
2010. in the presence of uncertainty via probabilistic
[12] Andronikos, T., “Classification of SPARQL automata”, Proceedings of the 3rd
queries into equivalence classes of relevant International Workshop on Algorithmic
queries”, International Journal of Advanced Aspects of Cloud Computing (ALGOCLOUD
Research in Computer Science, December 2017), 4-8 September, Vienna, Austria, 2017.
2017, Volume 8, No. 9, pages 152-159. [24] Papalitsas C., Giannakis K., Andronikos T.,
[13] Hua, M., Pei, J.: Probabilistic path queries in Theotokis D., Sifaleras A., “Initialization
road networks: traffic uncertainty aware path methods for the TSP with Time Windows
selection. In: Proceedings of the 13th using Variable Neighborhood Search”, 6th
International Conference on Extending International Conference on Information,
Database Technology, pp. 347–358, ACM Intelligence, Systems and Applications (IISA),
(2010). 6-8 July, Corfu, Greece, 2015.
[14] Andronikos T., Stefanidakis M., Papadakis I., [25] Papalitsas C., Andronikos T., Karakostas P.,
“Adding Temporal Dimension to Ontologies “Studying the impact of perturbation methods
via OWL Reification”, Proceedings of the on the efficiency of GVNS for the ATSP”,
13th Panhellenic Conference on Informatics - Proceedings of the 6th International
PCI 2009 Conference, 10-12 September, Conference on Variable Neighborhood
Corfu, Greece, IEEE Computer Society, pp. Search (ICVNS 2018), 4-7 October, Sithonia,
19-22, 2009. Halkidiki, Greece, 2018.
[15] Lian, X., Chen, L., Wang, G.: Quality-aware [26] Papalitsas C., Andronikos T., Karakostas P.,
subgraph matching over inconsistent “Studying the impact of perturbation methods
probabilistic graph databases. IEEE on the efficiency of GVNS for the ATSP”, 6th
Transactions on Knowledge and Data International Conference, ICVNS 2018,
Engineering 28(6), 1560–1574 (2016). Sithonia, Greece, October 4–7, 2018, Revised
[16] Reynolds, D.: Position paper: Uncertainty Selected Papers. Springer Theoretical
reasoning for Linked Data. In: Workshop, vol. Computer Science and General Issues,
14 (2014). Volume 11328, pp. 287-302, ISBN: 978-3-
[17] Schoenfisch, J.: Querying probabilistic 030-15842-2 (Print) 978-3-030-15843-9
ontologies with SPARQL, GI- (Online), 2019.
Edition/Proceedings 232, 2245–2256 (2014). [27] Papalitsas C., Andronikos T.,
[18] Krompaß, D., Nickel, M., Tresp, V.: “Unconventional GVNS for Solving the
Querying factorized probabilistic triple Garbage Collection Problem with Time
databases. In: International Semantic Web Windows”, (MDPI - Open Access
Conference, pp. 114–129. Springer (2014). Publishing), Technologies 2019, 7(3), 61;
[19] Khan, A., Chen, L.: On uncertain graphs https://doi.org/10.3390/technologies7030061.
modeling and queries, Proceedings of the [28] Papalitsas C., Karakostas P., Andronikos T.,
VLDB Endowment 8(12), 2042–2043 (2015). “A Performance Study of the Impact of

www.ijera.com DOI: 10.9790/9622-1011012330 29 | P a g e

Theodore Andronikos. International Journal of Engineering Research and Applications
www.ijera.com
ISSN: 2248-9622, Vol. 10, Issue 11, (Series-I) November 2020, pp. 23-30

Different Perturbation Methods on the

Efficiency of GVNS for Solving TSP”,
(MDPI - Open Access Publishing), Applied
System Innovation 2019, 2(4), 31;
https://doi.org/10.3390/asi2040031.
[29] Papalitsas C., Andronikos T., Giannakis K.,
Theocharopoulou G., Fanarioti S., “A QUBO
Model for the Traveling Salesman Problem
with Time Windows”, (MDPI - Open Access
Publishing), Algorithms 2019, 12(11), 224;
https://doi.org/10.3390/a12110224.
[30] Paz, A.: Introduction to probabilistic
automata, Academic Press, Inc., Orlando, FL,
USA (1971).
[31] Rabin, M.O.: Probabilistic automata,
Information and Control 6(3), 230–245
(1963).
[32] Cormen, T. H.; Leiserson, C. E.; Rivest, R. L.
& Stein, C. Introduction to Algorithms, Third
Edition, The MIT Press, 2009.

Theodore Andronikos*. “Complex Matrices for the Approximate Evaluation of Probabilistic

Queries.” International Journal of Engineering Research and Applications (IJERA), vol.10
(10), 2020, pp 23-30.

www.ijera.com DOI: 10.9790/9622-1011012330 30 | P a g e