of thesis works – Department of Software Systems and Technologies

List of topical topics Themes of dissertation research. Topics are grouped by supervisors. The student can choose the proposed topic or agree his own with the supervisor.

Bychkov Oleksii Serhiiovych

22 topics

(PN) - extension of HTN/strips/POMDP hierarchical schedulers for epistemically incompatible prerequisites

Development of planning algorithms for cases where the prerequisites for actions contain mutually incompatible information from different sources. Proof of completeness and correctness.

Planning under uncertainty

Federated learning in a multi-agent network of clinics to train medical models without sharing patient data

Development of FedAvg/FedProx protocol with privacy protection through differential privacy. Experimental verification on distributed datasets of radiographs.

Federated learning, privacy

Multi-task training for simultaneous detection of multiple pathologies and localization on medical images

Joint training of classification and segmentation for the detection of pathological zones on radiographs. Operability through Grad-CAM and derivative methods. Integration of explanations into the Dr.Case report.

Multi-task learning, XAI

Possibilistic Attention Mechanism in Transformer Architectures for Medical LLM

Replacing the usual softmax attention with (PN) -attentionwith a couple of features of possibility and necessity. Use in Dr.Case for the context of the patient's medical history.

Attention, LLM, transformers

Possibilistic Information Gain for optimal selection of clarifying questions in the diagnostic session

The development of the FocusedQuestionSelector of the Dr.Case system through the replacement of the classical Shannon entropy with a (PN)-measure of uncertainty. Theoretical substantiation of bonuses for specificity (×10/×5/×2 in v2.0) through the theory of possibilities.

Information theory, FocusedQuestionSelector

Self-supervised pre-training on unlabeled medical images for diagnostic DSS

Adaptation of SimCLR/MAE/DINO methods to the radiograph domain. Experiments showing the benefit of pre-trained models for rare diseases (<100 examples).

Self-supervised learning

Specificity-aware Bayesian update in multimodal medical DSS

Extension of the BayesianAnswerProcessor of the Dr.Case system (v2.0) to multimodal sources: symptoms, laboratory indicators, medical images. Calibration of likelihoods against real clinical data.

Bayesian inference, multimodal

Multi-agent architecture of intelligent medical diagnostics Dr.Case based on (PN)models

Extension of the Dr.Case monoagent system to a consilium of specialized diagnostic agents (cardiologist, neurologist, infectious disease specialist, etc.) with a protocol for coordinating diagnoses in (PN)formalism. Evidence of concilium convergence to correct diagnosis in the presence of pathognomonic symptoms.

Multi-agent diagnosis, Dr.Case

Detection of rare hematological diseases due to pathognomonic combinations of blood parameters

Extension of SymptomDrivenDiscovery of the Dr.Case system to laboratory data. Detection of pathognomonic combinations (for example, patterns in leukemia, aplastic anemia). Evaluation of sensitivity/specificity on data from real clinics.

Rare disease detection

Som + Focal-Loss NN Hybrid Architecture for Unbalanced Classes of Diseases: Theoretical Justification and Expansion

Evidence of TwoBranchNN_BMU_Focal architecture matching with guaranteed recall ≥ 99.5%. Extension to som hierarchy for rare diseases (<5 examples).

SOM, Focal Loss, deep learning

Deep neural networks for the analysis of chest radiographs with an assessment of epistemic uncertainty

Multi-label CNN architecture (based on ResNet/EfficientNet/Vision Transformer) for the detection of pneumonia, tuberculosis, COVID-19, neoplasms. Probability calibration through temperature scaling and Monte Carlo dropout. Integration as the 19th Dr.Case module.

Chest X-ray, CNN multi-label

Graphic neural networks for integrating laboratory parameters, symptoms and medical images into Dr.Case

Graphical representation of the medical case: nodes — symptoms, analyzes, images; ribs — clinical connections. Graph Neural Network for integrated diagnosis ranking.

Graph Neural Networks

Experimental calibration (PN)models for specific medical and engineering applications

Methodology for estimating parameters of capability/necessity function pairs from real clinical and engineering data. Benchmark v Bayes/Dempster-Shafer.

Calibration, validation

Epistemically persistent consensus protocols in swarm mobile networks with partial failures

Adaptation of Byzantine fault-tolerant consensus (PBFT, HotStuff) algorithms to conditions of high mobility and channel degradation. Theoretical assessment of the stability limit.

Byzantine consensus, MANET

Multi-agent reinforcement learning with formal epistemic failure

Extension of Marl algorithms (QMIX, MADDPG, MAPPO) by the mechanism of epistemic failure as a separate element of the scheduler output space. Experiments on UAV swarm simulators in electronic warfare conditions.

MARL, epistemic abstention

Resolving Dempster-Schafer paradoxes in multi-agent medical consiliums

In-depth analysis of 12 known paradoxes through the prism of (PN)-axiomatics. Constructive solutionof the 12th paradox (previously the school solved 11). Use in Dr.Case consilium versions.

Dempster-Shafer, paradoxes

Theory of epistemic refusal in medical expert systems: refusal of diagnosis in case of insufficient information

Formalization of the conditions under which the Dr.Case system should refuse to issue a diagnosis and transmit the case to the doctor. Integration into the Diagnosis Cycle as the fifth stopping criterion after Dominance/Stability/Need_Test/Safety.

Epistemic abstention, safety

Theory of functional stability of hybrid automata with (PN) -parameters for swarm control under uncertainty

Development of the theory of functional stability of the school (previous doctorate of the author) in the case of hybrid automata parametrized by pairs of functions of possibility/necessity.

Functional stability, hybrid automata

Three Layer Medical LLM Memory Architecture for Patient History

Working/semantic/episodic memory in transformer architectures. Integration with the Dr.Case NLP module for the context of patient visits.

Memory, LLM

Formal (PN)Theory of Five Layers of UAV Autonomous Swarm Architecture

Unified mathematical description of the reflex, reactive, deliberative, cooperative and social layers of swarm architecture in a single formalism (PN)models. Axiomatic justification of decision consistency between layers.

5-layer architecture, (PN)-models

Formal epistemic failure theory in neural network architectures and in large language models

Generalization of the concept of epistemic failure from a swarm of UAVs to general-purpose neural network architectures. Proof of the existence of a «safe» output-failure beyond resource constraints.

Neural abstention

Time models for the interpretation of complete blood count and biochemical blood test with prediction of the development of pathology

Transformer/Mamba based architecture for patient laboratory time series. Prediction of the dynamics of the disease, detection of latent pathologies by abnormalities in the CBC.

Time-series, CBC, biomarkers