Contact: Tamás Fischl
Level: MSc
Description: The student will implement a neural network capable of accurately and quickly detecting and classifying relevant objects (traffic participants) in aerial photographs. The labelled training set required to learn the neural network is available.

Contact: Tamás Fischl
Level: MSc
Description: Self-driving and driver assistance systems should be tested at different levels during development. Among others, functional testing (e.g. lane-keeping, automatic emergency braking) is necessary, where the behaviour of the test car is tested in predefined traffic situations, either on a test track or in road traffic. The student will be responsible for the implementation of a rule-based and/or machine learning-based software capable of aerially recording these traffic situations.

Contact: Tamás Fischl
Level: MSc
Description: While SAE Level 4 self-driving cars share the road with other human-driven cars, predicting the behaviour of cars (and pedestrians) in the vicinity of the self-driving car is of paramount importance. Graph Neural Networks (GNNs) can efficiently model traffic situations and have a negligible runtime requirement compared to convolutional neural networks, allowing them to run in real-time in self-driving cars. The aim of this thesis is to review examples from the graph neural-based literature, with a focus on behavior prediction, and to implement, test and optimize these examples on freely available data. A further objective is to test the best performing model(s) on realistic data and integrate them into the system.

Contact: Tamás Fischl
Level: MSc
Description: In industry, SWs are used almost everywhere in product or service development in manufacturing today, where they are developed by a team working closely together, even if they are developed by a team according to strict requirements. Over the years, a sufficient amount of data has been accumulated in terms of both specification and generated code to generate SW code based on a predefined specification. For example, there are numerous rule-based solutions for checking SW code, which perform the checking of a previously written code. Whether we use AI for code verification or code generation, the goal is to facilitate and assist human work. The scope of the task is far-reaching and resource-intensive, however, following on from the original idea, the aim of this thesis is to lay the foundations for such an application. It defines the level of detail of the specification, the architecture required, tests it by training on example data and verifies the results. The results obtained will be used to determine the direction of further development of the proposed method and, where appropriate, to provide feedback on the direction in which the specification used in industry should be changed in order to make future applications more successful and accurate.

Contact: Tamás Fischl
Level: MSc
Description: Industrial production processes (batch production) are carried out by automated production lines, mostly with minimal human intervention. Nowadays, however, a wealth of data is generated from such production lines, making production correctness, downtimes, interventions, bottlenecks, utilisation data available and traceable. This information, while allowing the tracking of the events of a series of production lines, does not allow a human to determine the optimum production. E.g. how long will it take to produce a given quantity of different types, designs, quantities, changeover times, etc. of a given product with a given production line capability over a given period? For such a complex process, optimization using an AI algorithm is the most feasible way. The objective of this task is to describe, understand and document a real industrial production line in such a way that the resulting information can be optimized using an AI algorithm. Although the duration of the thesis is probably not sufficient to develop the topic effectively for industry, the aim of the thesis is to show how the information gathered from the manufacturing process can be summarized in a way that can be used to teach a suitable AI algorithm. The results will open up the possibility to describe the process in detail. Finally, the algorithm can be used to make a recommendation on which products should be produced on which lines in order to maximise productivity.

Contact: Tamás Fischl
Level: MSc
Description: With the increasing role of SW-intensive features in modern vehicles, the architecture of the vehicle computing system is also transforming. SW features are moved from separate microcontrollers to central microprocessors with larger capacity. Embedded operating systems provide a certain level of independence among the various processes, which is essential to guarantee freedom from interference to each function. However, to meet higher safety and security standards, better isolation must be achieved. The goal of this project is to study available low-level hypervisor-based virtualization technology and implement a proof of concept based on open-source SW components where multiple heterogenous partitions can coexist on the same microprocessor with guaranteed integrity and fault isolation. The common core must ensure the secure execution environment and independence of each of the partitions, as well as allocation and access to the external interfaces and reliable means of inter-partition communication. The main tasks include:
- Study the requirements for safe and secure SW architectures when multiple functions are implemented on the same processor.
- Identify existing open-source and commercial SW frameworks and components available for a proof-of-concept implementation.
- Implement a demonstrator on a microprocessor-based system (Raspberry Pi or similar) with multiple partitions at distinct safety and security integrity levels.

Contact: Tamás Fischl
Level: MSc
Description: A gyalogos védő légzsák különféle szenzorokkal érzékeli a tárgyak becsapódását a kocsi elején. A szenzor adatok alapján egy algoritmus dönti el, hogy kinyisson a légzsák vagy sem. Az algoritmus fix, de paraméterekkel alakítható. A feladat az optimális paraméterek megtalálása törésteszteken kapott adatok alapján. Minden jármű típust törési teszteknek vetnek alá, melyek során felvételre kerülnek különféle becsapódások, melyekre szeretnénk kinyitni a légzsákot (tényleges baleseti helyzet), és melyekre nem (focilabda becsapódás, kátyú fölött áthaladás, kavics felütődés, stb). A beágyazott osztályozógépnek gyorsan és gyakran kell futnia, hogy időben meghozza a döntést, ugyanakkor csak kevés áramot fogyaszthat, és a HW költség sem lehet túl magas. Emiatt a beágyazott rendszerben nem használunk neurális hálókat, hanem egy egyedi döntési fa és logikai klasszifikáció elemeit ötvöző rendszerről van szó. Ugyanakkor a törésteszteken szerzett adatok feldolgozására és a paraméter kereső algoritmus kifejlesztésére offline van lehetőség, így az szinte bármilyen módszert és hardvert alkalmazhat. A feladat annak a kutatása, hogy lehetséges-e felügyelt tanulási módszerekkel (vagy egyéb technikával) a jelenlegi megoldáshoz képest számottevő javulást elérni, vagy átfogalmazni az optimalizációs feladatot úgy, hogy az a mostaninál lényegesen gyorsabban találja meg a globális optimumot.

Contact: Tamás Fischl
Level: MSc
Description: a PE oldali 0km reportokat kellene megvizsgálnia, ha valami nem jól alakul új pipeline részeket / funkciókat megírni pl. DataPreparation. Ez Python.
Ha van új vizsgálati metódust hoz akkor Andrással együtt műnödve bedolgozni a kódba.
Plusz van egy SW fejlesző csapatom ahonnan mindig vannak lepattanó munkát amivel kitöltenénk a maradék idejét.
Pl Dokupedea-ba kellene felvinni és bedolgozni strukturálisan pár lessions learned témát. ( Ez pedig embedded C. )
Ha tudunk találni olyat aki mind a kettőhöz konyít az lenne a legjobb, de ha választani kell akkor inkább C-s legyen ( python-ra könnyen betanul )

Contact: Tamás Fischl
Level: MSc
Description: Sensor validation testing is performed in different test chambers according to the customer requirements and international standards. The sensor data and the test bench data are monitored with diferent SW-s and stored on the local servers. Goal: Send and analyse test bench data synchronized with sensor data automatically on Hadoop server during sensor validation projects

Supervisor: László Gulyás
Description: Reinforcement learning is the machine learning technique behind the vast recent success of Artificial Intelligence applications. Multi-agent reinforcement learning (MARL) is the application of this technique to problems where multiple, independent, autonomous actors (decision makers) are present. These problems include two player games (e.g., tic-tac-toe, chess, go, etc.) or multi-player environments (e.g., starcraft -- to stay in the domain of games). The fundamental problem is the non-stationarity of the environment, since agents are part of each other’s world.

This thesis will explore the domain of MARL and study a selected algorithm by reimplementing one of the well-established methods from the literature and applying it to a chosen game. Both the game and the selected method is to be discussed with the supervisor and finalized at the submission of the thesis topic. The use of the Open Spiel platform (OpenSpiel: A Framework for Reinforcement Learning in Games) is encouraged.
Recommended level: MSc thesis
Recommended previous knowledge: C++ or Python

https://towardsdatascience.com/ive-been-thinking-about-multi-agent-reinforcement-learning-marl-and-you-probably-should-be-too-8f1e241606ac

https://www.dcsc.tudelft.nl/~bdeschutter/pub/rep/10_003.pdf

https://www.cs.utexas.edu/~larg/ijcai17_tutorial/multiagent_learning.pdf

Kaiqing Zhang, Zhuoran Yang, Tamer Başar: Multi-Agent Reinforcement Learning: A Selective Overview of Theories and Algorithms https://arxiv.org/pdf/1911.10635.pdf

https://github.com/deepmind/open_spiel

Supervisor: László Gulyás
Description: Despite their recent success, deep learning architectures suffer from important scalability issues, i.e. the actual artificial neural networks (ANN) become unmanageable as the number of features increases. Mocanu, Cavallaro, Liotta et al. recently proposed a new approach called sparse evolutionary training (SET), that the ANN training process is accelerated by enforcing various (sparse) network structures (i.e, scale freeness, small worldness).

The topic of this thesis is

• To replicate the seminal work of the above authors, and
• To extend it by applying the method to different application areas (e.g., different datasets).

Recommended level: MSc thesis (simplified version for BSc is possible)
Recommended previous knowledge: Lucia Cavallaro, Ovidiu Bagdasar, Pasquale De Meo, Giacomo Fiumara, Antonio Liotta: Artificial neural networks training acceleration through network science strategies https://www.researchgate.net/publication/344269018_Artificial_neural_networks_training_acceleration_through_network_science_strategies

Decebal Constantin Mocanu, Elena Mocanu, Peter Stone, Phuong H. Nguyen, Madelaine Gibescu, Antonio Liotta: Scalable Training of Artificial Neural Networks with Adaptive Sparse Connectivity inspire https://www.nature.com/articles/s41467-018-04316-3

 

Supervisor: László Gulyás
Description: Swarm intelligence is defined as a collective behavior of a decentralized or self-organized system. These systems consist of numerous individuals with limited intelligence interacting with each other based on simple principles. Although individuals are not very intelligent and there is no central leader or any strict hierarchical organization to dictate how individuals should behave, the interaction between individuals can help the system, as a whole, achieve intelligent behavior. To put it short: no individual has overall knowledge about the system or the task, everyone just carries out simple, local actions, but the community achieves a system level goal.
Stigmergy is a mechanism of indirect coordination, where the action of one agent changes the (physical) environment (i.e., leaves a trace) that influences (stimulates or inhibits) the behavior of another agent in the system. Stigmergetic coordination is often at the heart of the algorithms behind swarm intelligence.
The task of the thesis is to study the workings (i.e., the dynamic unfolding of the production of the solution) of a set of stigmergetic swarm intelligence algorithms by measuring the dynamics of the variance and entropy of the system (or the environment). The working hypothesis is that the time-evolution of these measures can characterise these algorithms well.
Recommended level: MSc thesis
Recommended previous knowledge: N.A.

Supervisor: László Gulyás
Description: RoboCup is an international scientific initiative with the goal to advance the state of the art of intelligent robots. To fulfill this goal RoboCup proposed several different competitions (leagues) in different domains. The RoboCup Soccer Simulation League (SSim) is one of the oldest leagues, focusing on artificial intelligence and team strategy. Independently moving software players (agents) play soccer on a virtual field inside a computer. There are 2 subleagues: 2D and 3D.
Our long term goal is to build the competences (team of students and researchers) necessary to participate in the regular international RoboSoccer Simulation competitions. We are currently at the first steps, setting up the team and infrastructure. Be prepared for intense learning, strategizing, but also for a lot of configuration tasks.
The thesis is to

• Review the state of the art of the RoboCup Soccer 2D Simulation League and its most successful algorithms.
• Create a base library that enables the easy implementation of certain classes of team algorithms and experimenting with them.

Recommended level: MSc thesis
Recommended previous knowledge: Robot Soccer (RoboCupSoccer Simulation) https://www.robocup.org/leagues/23, https://ssim.robocup.org/

 

Supervisor: László Gulyás
Description: Artificial Neural Networks (ANNs) are highly successful machine learning models and achieve human performance in various domains such as image processing and speech synthesis The density of connections within these artificially constructed networks is typically fairly high, despite various indications that sparse networks could yield better results.
Many automatic methods exist that yield sparsity in ANNs. Some are well-established regularisation methods (e.g. L1-regularisation), while others are novel, like Srinivas et al., Louizos et al., or Stier et al., etc.
The thesis will study a set of regularisation methods (with various parameters) and jointly assess their efficiency as well as the structure of the resulting network, looking for potential correlations between efficiency and network structure.
Recommended level: MSc thesis
Recommended previous knowledge:

• Suraj Srinivas, Akshayvarun Subramanya, and R Venkatesh Babu. 2017. Training sparse neural networks. In Computer Vision and Pattern, Recognition Workshops (CVPRW), 2017 IEEE Conference on. IEEE, 455–462.
• Christos Louizos, Max Welling, and Diederik P Kingma. 2017. Learning Sparse Neural Networks through L 0 Regularization. arXiv preprint, arXiv:1712.01312 (2017)
• Stier, Julian & Granitzer, Michael. (2019). Structural Analysis of Sparse Neural Networks. Procedia Computer Science. 159. 107-116. 10.1016/j.procs.2019.09.165.

 

Supervisor: László Gulyás
Description: Despite their recent success, deep learning architectures suffer from important scalability issues, i.e. the actual artificial neural networks (ANN) become unmanageable as the number of features increases. A recent approach to address the scalability problem of ANNs works with sparse neural networks, i.e., with network architectures that have a relatively low percentage of possible links present. Stier and Granitzer experimented with a set of specific low-density network structures and studied the accuracy of the resulting models, as well as the efficiency of training these networks, using the MNIST dataset.

This thesis will replicate (some of) the experiments of Stier and Granitzer and possibly extend them to other application domains (i.e., datasets).
Recommended level: MSc thesis (simplified version for BSc is possible)
Recommended previous knowledge: Julian Stier, Michael Granitzer: Structural Analysis of Sparse Neural Networks https://www.researchgate.net/publication/336542921_Structural_Analysis_of_Sparse_Neural_Networks

Supervisor: László Gulyás
Description: Despite their recent success, deep learning architectures suffer from important scalability issues, i.e. the actual artificial neural networks (ANN) become unmanageable as the number of features increases. A recent approach to address the scalability problem of ANNs works with sparse neural networks, i.e., with network architectures that have a relatively low percentage of possible links present. Hou and Kwok experimented with specific ANN architectures on specific datasets (MNIST, CIFAR-10, etc.) using a specific sparsification method. They studied the structure of the emerging trained neural networks and looked for evidence of a truncated power law degree distribution.

This thesis will replicate (some of) the experiments of Hou and Kwok and possibly extend them to other application domains (i.e., datasets) and/or to different sparsification methods and parameters. An important possible extension is to study to connection between the structure of the sparse network and its accuracy.
Recommended level: MSc thesis (simplified version for BSc is possible)
Recommended previous knowledge: Lu Hou, James T Kwok: Power Law in Sparsified Deep Neural Networks https://arxiv.org/abs/1805.01891

Supervisor: László Gulyás
Description: The direct route to artifacts (i.e., products, systems, etc.) is via design, i.e. by specification and subsequent realization (implementation). Evolutionary Technology takes an indirect route: the task is to produce an abundance of forms and functions of a practically endless variety by means of evolutionary methods. This implies the twin challenges of the ’arrow of complexity’ and ’open-ended evolution’, i.e. of producing increasingly complex machinery in the course of time, and doing that in a persistent, self-supporting process propelled by entirely endogenous causes.

Open-ended evolution is widely recognized as a difficult and unsolved problem. John Holland, the founder of Genetic Algorithms, once asked: ”Can we produce an existence-proof model, akin to von Neumann’s model of self-reproduction, that exhibits open-ended evolution, with increasing diversity and complexity?” Currently there is no accepted general evolutionary theory for the origin of complexity or the maintenance of evolutionary change.

In the 2000s Kampis and Gulyás published a series of papers based on the FATINT (fat interaction, i.e., phenotype plasticity) agent-based model. The task of this thesis work is the replication (re-implementation and re-validation) of the original FATINT model, preferably using the NetLogo agent-based simulation platform. (No previous knowledge of NetLogo is required, as it is very easy to learn.)
Recommended level: MSc thesis (simplified version for BSc is possible)
Recommended previous knowledge: https://www.researchgate.net/publication/228353897_Evolutionary_Technology_and_Phenotype_Plasticity_The_FATINT_System

https://www.researchgate.net/publication/5361726_Full_Body_The_Importance_of_the_Phenotype_in_Evolution

https://ccl.northwestern.edu/netlogo/

Supervisor: László Gulyás
Description: The topics below are to be refined via 1-on-1 discussions with the supervisor. Please, contact him for further information.

• Binary DCGAN on limited-sized datasets.
• DCGAN for line drawings on size-contrained datasets.
• Multi-agent reinforcement learning (MARL) in massively multi-agent systems
• Learning to swarm -- multi-agent reinforcement learning for swarm algorithms
• Connections between Schelling's Segregation model and Ant Sort (Cemetery Organisation by Ants)
• The Application of Machine Learning to Model Micro Behavior in Agent-Based Simulations
• Developing/Learning Fuzzy Rule Sets for Agent-Based Simulations

Current research interests: computational intelligence, cognitive robotics, memetic algorithms

Current research interests:societal intelligence, evolutionary algorithms, agent-based social simulations, social networks, intelligent systems

I am working with intelligent systems Keywords: human-computer-robot collaboration, cognitive science, neural systemns, target-oriented systems, distributed intelligence, educational gamesk

Current research interests:complex fluids, statistical physics