Examensarbeten och uppsatser / Final Theses

Framläggningar på IDA / Presentations at IDA

Se även framläggningar annonserade hos ISY och ITN i Norrköping / See also presentations announced at ISY and ITN in Norrköping (in Swedish)

If nothing is stated about the presentation language then the presentation is in Swedish.

• 2022-01-28 kl 08:15 i https://liu-se.zoom.us/j/66784186523?pwd=UEdKZU94K0xzTlRFeThUbGJhZzJxZz09
  
  Machine learning for optimal liquidation of financial assets
  
  Författare: Felix Kimiaei, Samuel Persson
  Opponenter: Benjamin Cajo, Viktor Gustafsson, John Lindmark
  Handledare: George Osipov
  Examinator: Marco Kuhlmann
  Nivå: Avancerad (30hp)
  
  The paper aims to evaluate the usefulness of the Deep Reinforcement Learning frameworks Deep Q-Network (DQN) and Proximal Policy Optimisation (PPO) when considered for an optimal execution problem. The models are evaluated in the context of execution horizons which are full trading days, as well as using some novel inputs. The data resolution is minute-wise, and the assets considered are US Nasdaq equities with high liquidity. The benchmark strategy used to evaluate the results is a Time Weighted Average Price (TWAP) strategy. The conclusion made is that both the DQN and the PPO framework in this paper can outperform a TWAP strategy with respect to the considered metrics. Furthermore, the results indicate that the inclusion of additional engineered features in the model input improves performance. Collecting the results, this paper indicates that both DQN and PPO can be useful in the time-restricted, one-sided, trading problem as formulated in this paper.
  
  
• 2022-01-28 kl 10:15 i https://liu-se.zoom.us/j/67900805047?pwd=TmwrS0k1a3ZCclBQQVBYbS9qSFNrZz09
  
  Prediction of User Actions with an Association Rules and two Neural Network Models in a Configuration Environment
  
  Författare: Hampus Elinder
  Opponent: Niklas Granander
  Handledare: John Tinnerholm
  Examinator: Lena Buffoni
  Nivå: Avancerad (30hp)
  
  Extensive configurators might suffer from latency due to their complexity - something which could hurt user experience and, in turn, sales. Machine learning models could potentially learn the patterns of user actions by looking at past configuration sessions. The models would then be able to predict the next action in a configuration session by looking at a few, past actions which can allow for pre-caching of data and, in the end, reduce latency. This thesis aims at testing the predictive capabilities of an Association Rules model and a Long Short-Term Memory and a Gated Recurrent Unit network. It found that all three models are capable of predicting future actions with the Gated Recurrent Unit producing the most accurate predictions.
  
  
• 2022-02-01 kl 10:15 i https://liu-se.zoom.us/j/65210771050?pwd=dnFNcmlnNHAzdVlhUnl6TUdSdXhBUT09
  
  Comparison of Distance Metrics for Trace Clustering in Process Mining – An Effort to Simplify Analysis of Usage Patterns in PACS
  
  Författare: Christoffer Sjöbergsson
  Opponent: Erik Mansén
  Handledare: Jonas Wallgren
  Examinator: Cyrille Berger
  Nivå: Avancerad (30hp)
  
  This study intended to validate if clustering could be used to simplify models generated with process mining. The intention was also to see if these clusters could suggest anything about user efficiency. To that end a new metric where devised, average mean duration deviation. This metric aimed to show if a trace was more or less efficient than a comparative trace. Since the intent was to find traces with similar characteristics the clustering was done with characteristic features instead of time efficiency features. The aim was to find a correlation between efficiency after the fact. A correlation with efficiency could not be found.
  
  
• 2022-02-08 kl 14:15 i https://liu-se.zoom.us/j/66611133278?pwd=OGFrNmFiSU5VdHMwMFJ3Rm8yaHI0dz09
  
  Evaluating the effectiveness of free rule sets for Snort
  
  Författare: Niklas Granberg
  Opponent: Sofia Gyulai
  Handledare: Mohammad Borhani
  Examinator: Andrei Gurtov
  Nivå: Avancerad (30hp)
  
  As the use of the internet increases in the daily life of people and larger entities, the threats that comes with that connection also grows. Attacks happen all the time and many have serious consequences that disrupts the daily processes and cost massive amounts of money. To fight back against these attacks, defensive tools are used to find and counter attacks. One of these tools is Snort. Snort find malicious data packets and warns the user. Snort relies on a list of signatures of different attacks, with a few of these rule sets being free. But how well can Snort defend a network using these free rule sets? Using a testing network and filling it with realistic background traffic, a series of rule sets are evaluated against a set of common attacks and tools. The performance hit when defending in a high speed, high bandwidth environment is evaluated as well. The end results show favour to the Emerging Threats rule set, with the warning that the rule set did not find 100% of the involved attacks. As for performance, Snort could not handle the most extreme amounts of traffic, with the rate of dropped packets making security dubious, but that occurred at the absolute peak of what consumer hardware can provide.
  
  

Page responsible: Ola Leifler
Last updated: 2020-06-11