=Paper=
{{Paper
|id=Vol-3816/paper69
|storemode=property
|title=Is the Rule-based Order in AI Dead, or Is Still Kicking?
|pdfUrl=https://ceur-ws.org/Vol-3816/paper69.pdf
|volume=Vol-3816
|authors=Vassil Vassilev
|dblpUrl=https://dblp.org/rec/conf/rulemlrr/Vassilev24
}}
==Is the Rule-based Order in AI Dead, or Is Still Kicking?==
https://ceur-ws.org/Vol-3816/paper69.pdf
Is the Rule-based Order in AI Dead, or Is Still Kicking?⋆
Vassil Vassilev1,∗
1 Sofia University, 5 James Bourchier Blvd., 1164 Sofia, Bulgaria and London Metropolitan University, 166-220 Holloway
Road, N7 8DB, London, UK
Abstract
This article focuses on the possibility of revival of the classical rule-based approach in AI for building
heterogeneous AI systems. It is based on personal experience while working on several research and
innovation projects in diverse areas - business process management, unauthorized intrusion detection,
malicious interference protection, digital forensics and diagnostics, data management, etc. Despite
their differences, all these areas share something in common: they require multiple operations to be
executed in a single transaction and incorporate heuristic rules for different purposes, related to data,
knowledge, and operation management. The position of this article is that combining the classical rule-
based approach from the early days of AI with more recent developments in AI, such as data-focused
machine learning and utility-based reinforcement learning, as well as utilization of the recent
technological developments on the cloud and at data centers can be beneficial for widening the real-
world application of AI. The challenges, which this complexity creates require joined efforts of
academic researchers, industrial engineers, and business enablers. Collaboration between them across
the board can be highly beneficial and the author is looking for opportunities in this direction.
Keywords
Heterogeneous AI; Knowledge-based Systems; Semantic Technologies; Heuristic Rules; Integration,
Synchronization and Automation
1
1. Introduction
The history of AI witnessed several shifts in its dominant paradigm - from the initial amazement
of the movements of digital amoebas to the decision to embed our knowledge in them to make
them more active to the relief of letting them learn themselves to the complete relying on the
unknown intelligence exhibited by chatting boxes... Despite these twists and turns, one baseline
remained unchanged - the rule-based order still rules in AI!
This paper reflects on the experience of several AI-based projects from recent years
completed under the eye of the author at GATE Institute of Sofia University and the Cyber
Security Research Centre of London Metropolitan University. They all have something in
common - in one way or another, they rely on rules.
The different paradigms adopted in these projects solve different tasks - in decision-making
for choosing alternatives for continuation of the operations, in planning the activities for
achieving the goals, in controlling the execution of the operations to stay on track, in learning
from the environment to improve the outcomes, in looking back at the experience to improve
the planning and in explaining the results to bring confidence in the solutions ... This easily leads
us to the belief that the way forward is the hybridization and the practical way of uniting multiple
RuleML+RR2024: 8th International Joint Conference on Rules and Reasoning, September 16–18, 2024, Bucharest,
Romania
∗ Corresponding author.
vassil.vassilev@gate-ai.eu
0000-0003-4361-4830
© 2024 Copyright for this paper by its author. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
paradigms to achieve this is through the use of rules. Of course, this is a complex undertaking,
but it does not need to be chaotic.
The paper analyses the experience in hybridization within several projects completed or still
underway in Sofia and in London, executed in collaboration between the Cyber Security Research
Centre of London Metropolitan University and the GATE Institute of Sofia University. After a brief
description of several projects, showing the place of heuristic knowledge in them the paper
discusses three fundamental models of heuristic rules, which can meet the requirements of many
similar projects. At the end of the paper, the author focuses on three main directions of interest
for research and technological development, which are worth investigating further and can be
considered an invitation for collaboration.
2. Recent Project Experience
Our more recent projects which involve heuristics are as follows:
Assessing the Logical Vulnerability of Transactional Systems: This was the last of a series of projects
funded by Lloyds Banking System completed at the Cyber Security Research Centre of London
Metropolitan University during 2018-2019 [8]. It aimed at incorporating security threat
intelligence into a logical model of financial transaction processing under duress and analyzing
the logical vulnerability of the security policies at the design phase. The most important
achievement of the project was the development of a method for modeling and analysis of
transaction processing, based on combining general-purpose ontology of transactions under
security threats with an efficient algorithm for analysis of the security policies which utilizes
semantic indexing. In this project several types of rules were utilized: rules for choosing suitable
methods for detection, rules for imposing security policies for counteracting security threats, and
rules for analyzing the logical vulnerability of security policies.
Threat Intelligence for Unauthorized Intrusion Detection: This project was initiated in the autumn of
2019 by the Cyber Security Research Centre of London Metropolitan University in collaboration
with a local company that specializes in managing data centers for commercial businesses using
IaaS services provided by AWS public cloud. The objective of the original project was to perform
security analytics in real time using data from the company's customers' networks on the cloud.
Due to the outburst of the pandemic though, the original project was substantially modified.
Thanks to GATE Institute of Sofia University, which provided additional resources the project was
revived and continued as a joint project until its completion in 2020 [6]. The security data was
generated by simulation in London and sent for processing to Sofia in real-time, where it was
analyzed – first in real-time, using statistic methods of correlation, and later on offline, using
machine learning algorithms for forensic investigation of the accumulated data. The use of
heuristic rules in the project was linked to the orchestration of data management operations along
the pipeline for data processing from the moment the data was ingested on the cloud server up to
the moment its analysis was concluded and it was accumulated in one of the databases of the
cloud server. This project created the prototype of the future GATE Data Platform which in its
third enterprise version currently serves the needs of data processing of GATE Institute.
Risk Assessment in Transactional Systems: This project was coined by Lloyds way back in 2019, based
on the success of the previous projects, and started as an internal project at the Cyber Security
Research Centre of London Metropolitan University, but due to the outbreak of the pandemic, it
was completed as a collaboration project at GATE Institute of Sofia University in 2020 [5]. This
collaboration leverages combining the technological experience of the Cyber Security Research
Centre in ontological modeling of transactional systems with the experience of GATE Institute in
decision-making and stochastic optimization. Jointly we developed a new method for modeling
transactions under duress by adopting the Partially Observable Markov Decision Process model
of stochastic planning (POMDP). Instead of directly solving the problem using standard iterative
approximation, though, by restricting the model to the intelligence graphs developed for
vulnerability analysis we were able to solve the problem using an efficient recurrent algorithm.
Although the heuristic rules were embedded directly in the ontology of the transaction process
itself, lots of heuristics were applied manually to configure the model and to interpret the results
of the computation. This project currently continues as an innovation project under the Cyber
ASAP scheme of Innovate UK with potential commercialization aimed at automation of the default
configuration of the model during the modeling phase.
Air Pollution Monitoring and Environment Factors Analysis: This project started in Sofia in 2021, at
that time considered to be the most polluted capital in the EU. It served as the first pilot
application of the prototype of GATE Data Platform for the provision of comprehensive
information in real-time about the state of air pollution in the city using information obtained
directly from the monitoring stations across Sofia [4]. From a technological perspective, the most
interesting thing was the combination of real-time and offline data processing of both static data,
retrieved from a model of the urban infrastructure in ontological format, with external dynamic
data, coming from the sensor stations across the city in real-time, and the internal event data,
resulting from the interactive operations on the system interface. This concept was later on re-
deployed by the same team in 2022 in London, where additional static data about the medical
prescription inside the catchment area was used to correlate the local air pollution and
respiratory diseases [3]. In both cases, a bunch of heuristics was used to integrate, synchronize,
orchestrate, and interpret the data, coming from different data sources.
In addition to the above projects, in which the heuristics play an operational role and can be used
directly for the automation of various analytical tasks, we have been working on some projects,
in which the rule-based approach can have a more methodological role. This includes the internal
project for developing GATE Data Platform, the EU DiverSea project for developing an integrated
architecture for analysis of the maritime biodiversity of EU coastal seas, the DIANA CoDe project
for countering disinformation in media space, etc. The common theme in all these projects is the
orchestration of the data management work lows, which allows automation by applying a variety
of methods for data processing on the data platform, such as data uni ication, semantic data
enrichment, data low synchronization, and operation orchestration
3. Rule Design Considerations
In this section, we will consider three different approaches for modeling transactional systems
which differ substantially in their complexity but can be used as a baseline in applications in
which there is a need to account for domain-specific heuristics. They have been major research
themes in AI and Computer Science for some time on both theoretical and technological levels,
but the complexity of the problems leaves plenty of space for further investigations. At the same
time, they illustrate the different levels of depth in representing heuristic knowledge and its
impact on the overall system architecture. Their choice reflects the need to incorporate rules in
our projects, and our expertise and can be considered an invitation for further research and
collaboration.
3.1. Rules in Classical State-space Model of the Transactions
Figure 1: Inventory for Formulating Structural and Parametric Deterministic Rules.
The simplest template for formulating policy rules can be constructed in the classical state-space
model of dynamic systems, which was introduced in AI in the situation calculus [11]. It adopts a
functional view of the actions as mappers of the global states, infamously and misleadingly called
situations. As a sequence of actions, the formal model of the transactions in this approach is a
directed graph, leading from the initial state to a commit state - in the case of success, or, to a
rollback state instead - in the case of failure. In Fig. 1 we show a fragment of such a transaction
graph, focused on a single node representing a particular state, together with the related nodes.
On top of such a model, we can formulate multiple heuristic rules (policies) to support the use of
the transaction model for various purposes – design, monitoring, synchronization, orchestration,
automation, etc.
The rules can be formulated as functional dependencies of the results of the actions on the
values of the state parameters. potentially also accounting on the history:
Rule01: S1=f1 (D01S0=V11,D02S0=V12,…,D0kS0=V1k, Sm)
Rule02: S2=f2 (D01S0=V21,D02S0=V22,…,D0kS0=V2k, Sm)
...
Rule0n: Sn=fn (D01S0=Vn1,D02S0=Vn2,…,D0kS0=Vnk, Sm)
where fi are the transition functions corresponding to the potential actions that change the
current state. The rules are usually formulated in a logical format, which allows to combine
explicitly AND and OR conditions on the states. The security policies in such a case can be
interpreted conveniently as an AND/OR graph.
This template is basic – it does not represent the types of state descriptors, it does not account
for potential classifications, and it cannot distinguish synchronous and asynchronous activities
in the same model. This limits its use to static tasks during the design phase, which considers
only synchronous or asynchronous activities, but it is not suitable for operational control and
automation in real-time applications. From a purely theoretical perspective, it also faces several
principal problems, such as qualification, ramification, and frame problems [11], which may lead
to practical complications.
At the same time, this template is universal and can be helpful at the initial design, linked to
the structural configuration of the workflows and their parametrization. We have investigated a
series of structural patterns, providing sufficiently informative ground for specifying structural
heuristics concerning the composing and controlling of the workflows. To increase the efficiency
of the rule-based systems that use deterministic rules in this format, we also extensively indexed
the rules against the states and their parameters. Since this can be done entirely incrementally,
it can be very useful for business process management [10]. This approach is also suitable for
automation of the scheduling of data processing pipelines on general-purpose data platforms
since it does not account for fine-grained domain-specific knowledge which requires more
complex heuristics and can be easily mapped to the tools used to control the data processing
pipelines, such as AirFlow we used [1].
3.2. Rules in Situation-based Model of the Transactions
Figure 2: Inventory for Formulating Static and Dynamic but Deterministic Rules
We adopted the situational view of the actions a long time ago, but the practical value of this
approach came out after we managed to formulate the transactional model in terms of
Description Logic (DL) [14] and to represent it in serialized form as an ontology using standard
languages of the semantic technologies multi-layered cake - RDF/RDFS and OWL [15]. For this
purpose, we introduced three separate vocabularies into the description logic with appropriate
axiomatization of the theory:
- modeling the situations as DL concepts (interpreted semantically as a set and
represented in the ontology as OWL classes)
- modeling the actions to be DL properties (interpreted semantically as relations between
the situation sets and represented in the ontology as OWL properties)
- introducing events in the DL theory as another type of DL concept to model the
asynchronous activities
This approach has substantial advantages over the state space-based approach. Firstly, by
modeling the actions as relations between the situations we avoid some of the hurdles of the
state-space approach since the description of the situations does not need to be exhaustive,
which neutralizes the qualification problem. Secondly, utilizing the possibility of using
conceptual types in DL it is possible to distinguish static and dynamic concepts explicitly and this
way, to model both synchronous and asynchronous activities as independent. This conveniently
supports event-driven control and real-time operation in transactional systems. Thirdly, by
removing the explicit syntactic parametrization of the actions, which are no longer functions of
their own parameters but conceptual relations between situations we can adopt the implicit
semantic binding of their parameters to the parameters of the situations on the meta-level, which
resolves the frame problem in an unexpected and elegant way [9]. Finally, the possibility to use
taxonomies in DL theories and the corresponding serialized ontologies for both concepts and
properties allows us to employ object orientation in modeling and implementation as well.
Since the semantic cake maps the interpretation of different serialized representations onto
the same semantic domain, the interpretation of rules becomes semantically consistent with the
interpretation of the ontology itself and can be easily integrated with it. The standard way of
doing this is by using SWRL as a modeling language [15]. This allows to turn the models of
transactional systems into intelligence graphs, which embed data, facts, conceptual and heuristic
knowledge in a single repository, similarly to the knowledge graphs, which combine data, facts
and conceptual knowledge.. This approach allows to incorporate much richer heuristics
knowledge through the use of a separate domain ontology. We successfully used this approach
to analyze the logical vulnerability of security policies in banking, formulated as SWRL rules on
top of the ontology [8] and for combining domain-specific and problem-specific knowledge to
generate a more informative presentation of geolocation information [3-5]. It has the potential
to address the problem of explanation in AI through the use of a separate ontology of causality,
where a simple black-box approach can bring to the surface a deeper explanation based on the
causal relation between events, actions, and situations participating in the workflow of data
processing operations [7]. This methodology is currently under development at GATE Institute
within the framework of the EU Horizon DiverSea project.
3.3. Rules in Stochastic Models of the Transactions
Figure 3: Inventory for Formulating Rules in Non-deterministic Setting
Up to the moment we have considered only deterministic templates, which allow us to rely on
purely logical methods for modeling the heuristics as rules. However, in many practical
applications, it is impossible to formulate the rules precisely due to the impossibility of assessing
the environmental conditions or the subjective expertise with certainty and precision. Such is
the case in many typically analytical tasks which require assessing the risks, associated with the
normal functioning of transactional systems due to a variety of factors, such as unauthorized
intrusion, malicious intervention, unexpected faults, imprecise measurements, or insufficient
trustworthiness. In several of our projects in both the UK and Bulgaria which were focused on
controlling financial transactions, we tried to overcome the limitation of the deterministic
models by associating the degree of probability directly in the heuristic rules, considering the
transactions as Markov processes [5]. Fig. 3 shows a fragment of a graph that models
transactions under stress as a Partially Observable Markov Decision Process (POMDP) [16].
This model introduces non-determinism in two dimensions of the transactions – the
transition between situations (labeled using pij as a probability of transitioning from situation Si
to situation Sj) and the observations of the events in each situation (labeled using qij as a
probability of observing event Ej in situation Si). The non-deterministic transition in this case is
interpreted as caused by the conditions within the state, which can be both anticipated, but
unknown (i.e., results of tests), or unexpected, but recoverable (like security threats, malicious
interferences, device malfunctioning, etc.). The non-determinism of the observation is
interpreted as either probable appearance of the event (in the case of unexpected), or probability
for detection (in the case of anticipated).
Unfortunately, introducing some degree of expectation of the possible transitions, such as
probability for choosing an alternative route for continuing the transactions, and assigning a
degree of imprecision of the observations, such as probability for detection of the events along
the transactions, introduces additional difficulties for using this transactional model. Apart from
the computational complexity of the algorithms, used for evaluating global characteristics of the
transactions which typically require approximate solutions of Bellman-type of equations, the
additional difficulty comes from the dependence of these probabilities. While the first problem
in some cases can be solved by reduction of the original POMDP problem to an MDP problem,
which has precise solution using an efficient recurrent algorithm as we have shown in [5], the
second problem remains completely outside the mathematical brilliance. At the same time, this
creates a whole new world of opportunities for engineering pragmatism.
We have used this model successfully for quantitative assessment of the risks in online
transactions under security threats and for analyzing the impact of the precision of intrusion
detection for the purpose of designing security infrastructure with guaranteed low risks [2,6]. In
our most recent project funded by UK Innovate UK project (CyDRA), we have adopted this model
for the design of the system architecture of a commercial software product we are currently
developing specifically for vulnerability analysis and security risk assessment of cyber systems.
4. Opportunities for Further Development of the Rule-based Approach
From the above considerations, it is clear that the problem for modeling, controlling, and
analyzing the data processing in transactional systems is far from simple. But it also creates a lot
of opportunities for further development of the technologies and creating the methodological
basis for creating heterogeneous and distributed AI systems. While there is no chance to create
a generic solution that is universally applicable to all problems requiring intelligence, it is also
obvious that hybridization is the way forward. Based on such an understanding in this section
we will discuss the opportunities for achieving this from both the conceptual and technical side.
4.1. Extending the Rule-based Systems with External Ontologies
Knowledge is power but knowing the problem is only one of the conditions for finding
solution of complex problems. The ontological models embody domain expertise which can be
brough by the external experts to the software and data engineers to contextualize the solutions,
make the AI models more informative, and the algorithms for problem solving more efficient.
This direction of research is known as ontological logic programming (OLP). It promises to go
beyond the classical logic programming paradigm. To avoid some of the difficulties of adopting
it in practice we should balance better between theoretical generality and practical applicability.
Our own direction of research to achieve such a balance is the use of templates in the modeling
language and the use of indexing mechanisms in the heuristic rule-based inference.
4.2. Adding Heuristics for Setting the Probabilities in Stochastic Models
One of the practical difficulties in adopting more powerful stochastic models for solving real-life
problems involving incomplete, imprecise, and fuzzy knowledge is the setting of default
parameters of the stochastic models. The mathematical models behind POMDP and MDP
consider the prior probabilities as independent, while in reality, they are subject to logical,
temporal, and causal dependencies between the situations, events, and actions. By accounting
for these dependencies more realistic distribution of the probabilities can be set up. This is the
subject of a separate analysis which can lead to the adoption of logically consistent heuristics for
parametrization of the model at the design stage. We are currently working on a set of templates
for contextualizing the menu-driven interface of the transaction modeler, which makes use of
such dependencies. This would allow implicit accounting of domain-independent and even
domain-specific heuristics directly during the modeling process, which would increase the
quality of the model without the need of complex modeling experience.
4.3. Utilization of Rule-based System Architectures for Explanation
The adoption of a multi-layer architecture for data processing in which the data management,
the data analysis, and the explanation of the results appear on different levels requires a complex
explanation, that combines the logic of the separate levels. However, the explanation has its own
logic going back to the philosophical studies of scientific explanation by Carnap, Quine, Hempel,
and others, and although it is not realistic to expect full-depth coverage of this phenomena, the
ontology of causal dependencies between conditions, events, actions, and effects can be
developed even without philosophical depth, purely from technical common sense. Such an
ontology can be used for a black-box type of explanation generation which can address many
concerns in contemporary literature, linked to some hard legal issues of adoption of AI and the
need for developing Explainable AI.
5. Conclusion
So, the conclusion is simple: the rule-based approach from the early days of AI is not dead, and
the heuristic rules are still kicking in multiple places; even more – the rule-based systems are
increasingly more important to handle the real-life complexity of digital reality.
The projects considered in this paper have a wide scope: from purely security issues (fraud
detection, unauthorized intrusion detection, vulnerability analysis, and risk assessment), to
environment issues (air pollution monitoring, biodiversity analysis), and to their wider social
impact (in healthcare, business management and legal practice). However, all of the projects
have something in common - they involve complex transaction management which combines
multiple paradigms for data processing. Such a system can be orchestrated by a centralized
system, which requires system policies based on rules. The common denominators here are two:
knowledge modeling, which can be a basis for formulating and applying both domain-specific
and problem-specific policy rules, and the format of the rules themselves, which affects the
algorithms for data processing.
From the perspective of contemporary technological advancements, the adoption of the
above principles requires the adoption of a suitable platform for data processing, which utilizes
virtualization, containerization, and orchestration of software services in a cloud environment.
The recent shift of attention to AI also opens a wide horizon for the automatic configuration of
software services, control of the operational pipelines, layered visualization, and causal
explanation of the results on such a platform. This leads to truly heterogeneous and distributed
systems.
We have successfully implemented elements of such a complex solution at the two research
centers in London and Sofia, proving its viability. Although we are committed to continue
working after the above principles within our own centers, we are also very keen at collaborating
with other research groups sharing similar views. Particular potential for this exists in two
current projects (DiverSea and CyDRA), The large scope of DiverSea project, which covers most
of the shores of Europe, together with the wide representation of partners creates an excellent
opportunity for regional follow-up projects to apply and deepen the methodology for
investigating the biodiversity on a regional scale. On the other hand, the potential application of
CyDRA software currently under development creates an opportunity for direct use of the
models, methods and algorithms in other application domains where the focus is on
transactional information processing under various factors of risks. Particularly interesting is
the possibility for a follow-up project in the domain of healthcare, where the assessment of the
risks of developing certain diseases can substitute for the second opinion of the medical
professionals. Another interesting option here is the possibility to use the vulnerability analysis
and risk assessment for designing production lines with guaranteed safety of operation, which
is critical not only in manufacturing but also in food production.
Acknowledgements
The above considerations have been in the making for nearly a decade at two different institutions – the
Cyber Security Research Centre of London Metropolitan University and GATE Institute of Sofia University.
The projects referred to here have been funded from multiple sources in the UK (Lloyds Bank, Higher
Education Investment Fund, and Innovate UK, specifically the CyDRA project under CyberASAP program2),
the EU (Horizon 2020 WIDESPREAD Programme, Horizon Europe 2023 Programme, specifically the
DiverSea project under BIODIV Programme3), and Bulgaria (BG Government Operational Programme
Science and Education for Smart Growth). The author is grateful to all these organizations for their
continuing support and trust. However, all views expressed in the paper are of the author and in no way
reflect the official positions of these organizations on the problems discussed here.
During these years I was lucky to have excellent students. Many of them later became my PhD students
and project collaborators, like Pawel Gasiorowski, Viktor Sowinski-Mydlarz, Kanana Ezekiel, Karolina
Bataytite, Artur Nascicionis, Khalid Mohamed, Sorin Radu, Hristo Hristov, Martin Hristev, Dion
Mariyanayagam, Sabin Nakarmi and Reza Baghaeishiva, to mention some of them. In addition to the
understanding and the support from colleagues and friends at Sofia University and London Metropolitan
University, working with my students was a decisive factor in completion of the projects. Special thanks
to my old friend Doncho Donchev from GATE Institute, who created the algorithm for reducing the original
POMDP problem to a more tractable MDP problem.
2 Innovate UK, CyDRA: Risk and Vulnerability Assessment in Transactions under Security Threats, Cyber Security
Academic Startup Accelerator Program, Project 10139273 (2024).
3 EU Executive Agency, DiverSea: Integrated Observation, Monitoring and Prediction Architecture for Functional
Biodiversity of Coastal Seas, HORIZON-CL6-2022-BIODIV-01 Programme, Project 101082004 (2022).
A final word about the role of industrial partnership. The academic environment provides the working
environment for conducting research and technological advancement, but it is too sterile and often
prevents from facing the reality of real-world conditions. The author is grateful to many people from the
industry, who helped to overcome this constraint, but first and foremost, I would like to mention Tony
Phipps and Matt Lane, with whom I had many hours of interesting, inspiring, and productive discussions.
References
[1] V. Vassilev, V. Sowinski-Mydlarz, P. Gasiorowski, et al., Building A Big Data Platform using Software
without License Costs, in: L. Castro (ed.), Open Source Horizons, London: IntechOpen, 2024, pp. 29-52.
doi:10.5772/intechopen.1003268
[2] D. Dunsin, M. Ghanem, K. Ouazzane, and V. Vassilev, A Comprehensive Analysis of the Role of AI and
Machine Learning in Modern, Digital Forensics and Incident Response, Forensic Science
International, Volume 48, March 2024, Elsevier, pp. 1-22. doi: 10.1016/j.fsidi.2023.301675
[3] V. Vassilev, B. Virdee, K. Ouazzane, et al., Data Platform and Urban Data Services on Private Cloud, in:
Senjyu, T., So-In, C., Joshi, A. (eds.), Smart Trends in Computing and Communications, Springer LNNS,
Vol. 650, 2023, pp. 263-275. doi: 10.1007/978-981-99-0838-7_23
[4] V. Vassilev, V. Sowinski-Mydlarz, D. Petrova-Antonova, et al., Towards first urban data space in
Bulgaria, in: Proc. 8th Int. Smart Cities Conference (ISC2), Paphos, Cyprus, IEEE, 2022, pp. 1-7. doi:
10.1109/ISC255366.2022.9922237
[5] V. Vassilev, D. Donchev, and D. Tonchev, Risk Assessment in Transactions under Threat as a Partially
Observable Markov Decision Process, in: L. Amorosi et. Al. (eds.), Optimization in Artificial Intelligence
and Decision Sciences, Springer, 2022, pp. 50-66. doi: 10.1007/978-3-030-95380-5_18
[6] V. Vassilev, K. Ouazzane, V. Sowinski-Mydlarz, et al., Network Security Analytics on the Cloud, in: Proc.
13th Int. Conf. on Cloud Computing, Data Science & Engineering (Confluence), IEEE, 2023, pp. 151-
156. doi:10.1109/Confluence56041.2023.10048889.
[7] V. Vassilev, S. Ilieva, V. Sowinski-Mydlarz, et al., AI-based Hybrid Data Platforms, in: Curry, E., Scerri,
S. and Tuikka, T. (eds.), Data Spaces: Design, Deployment, and Future Directions, Springer, 2022, pp.
147-170. doi: 10.1007/978-3-030-98636-0_8
[8] V. Vassilev, V. Sowinski-Mydlarz, P. Gasiorowski, et al., Intelligence Graphs for Threat Intelligence and
Security Policy Validation of Cyber Systems, in: P. Bansal, et al. (eds.), Advances in Intelligent Systems
and Computing, Vol. 1164, Springer, 2020, pp. 125-139. doi: 10.1007/978-981-15-4992-2_13
[9] K. Bataityte, V. Vassilev, and O. Gill, Ontological Foundations of Modelling Security Policies for Logical
Analysis, in: I. Maglogiannis, et al. (eds.), IFIP Advances in Information and Communication Technology,
Vol. 583, Springer, 2020, pp. 368-380. doi: 10.1007/978-3-030-49161-1_31
[10] K. Ezekiel, V. Vassilev, K. Ouazzane, et al., Adaptive business rules framework for workflow
management. Business Process Management Journal, 2018: 25 (5), pp. 948-971. doi: 10.1108/BPMJ-
08-2017-0219
[11] J. McCarthy and P. Hayes, Some philosophical problems from the standpoint of artificial intelligence,
In: B. Meltzer and D. Michie (eds.), Machine Intelligence 4, Edinburgh University Press (1969), pp.
463–502. ISBN: 9780852240625
[12] J. Hintikka, Knowledge and Belief: An Introduction to the Logic of the Two Notions. King's College
Publications (2005). ISBN: 9781904987086
[13] J. Barwise and J. Perry, Situations and Attitudes. MIT Press (1983); ISBN: 9780262021890
[14] F. Baader, I. Horrocks, C. Lutz, and U. Sattler, An Introduction to Description Logic. Cambridge
University Press (2017). ISBN: 9781139025355
[15] P. Hitzler, M. Krötzsch, and S. Rudolph, Foundations of Semantic Web Technologies. Chapman &
Hall/CRC (2009). ISBN: 9781420090505
[16] Mausam and A. Kolobov, Planning with Markov Decision Processes: An AI Perspective. Morgan Claypool
Publishers (2012). ISBN: 9781608458868