|
|
Processing of time tables
Mrkus, František ; Fink, Jiří (advisor) ; Pilát, Martin (referee)
A goal of this thesis is to create an open-source application which could serve as foundation for public bus transport analysis and organizing, while di- rectly operating with timetables in a JDF format for a comfortable workflow. The application is centered aroud bus scheduling for public transport orga- nizers and agencies,including related functions such as displaying timetable sheets and departure/arrival lists, map visualization of the planned routes, and creation of custom timetables. All of these features were sucesfully im- plemented and tested on real-world data.
|
|
|
Procedural generation of levels for a realtime stealth game
Sedlák, Filip ; Černý, Vojtěch (advisor) ; Pilát, Martin (referee)
Levels in realtime stealth games are often tightly interconnected struc- tures with unique challanges and lock & key puzzles. However, there are no well known instances that attempt to generate these levels procedurally. We implement a small generic 3D realtime stealth game and a level genera- tion algorithm for it. Our game is composed of mechanics commonly found in most stealth centric games. Our generated levels resemble levels from modern stealth games in complexity and interconnectedness. They contain unique challenges for the player. Some generated level sections have short- comings, but are always playable. In summary, we believe our algorithm succeeded as a proof of concept and can be used in actual stealth games with additional content. Moreover, we contributed a concrete implementation of the cyclic generation algorithm, where the original source is vague on imple- mentation details. Our algorithm can be used to generate levels for stealth games, RPGs, and other genres that make use of lock & key puzzles.
|
|
|
Genetic programming methods for classification
Nagy, Marek ; Neruda, Roman (advisor) ; Pilát, Martin (referee)
This thesis examines using different genetic programming encodings and analyses if they can be used for classification machine learning tasks. We introduce Evolutionary Algorithms and their basic concepts as well as define our specific branch Genetic Pro- gramming that is used to generate formulas instead of a differently encoded parametric solution. We introduce Cartesian and Tree-Based encodings and operations on them needed for the algorithm to function properly. The proposed algorithms are implemented and their performance is tested and their results compared on multiple datasets. We then describe how to build and run our solution and discuss the results of the experiments. 1
|
|
|
Leveraging lower fidelity proxies for neural network based NAS predictors
Mintál, Samuel ; Neruda, Roman (advisor) ; Pilát, Martin (referee)
The performance of a neural network is dependent on several factors including its underlying architecture. The field of neural architecture search (NAS) is an important part of automated machine learning (AutoML) as it focuses on automatization of a previ- ously manually performed search process for the best performing architecture for a given task. Estimation of performances of architectures is an inseparable part of NAS. As the standard full training and consequent evaluation of architectures is computationally infeasible a lot of research is focused on creating less computationally demanding ways for performance estimation. In this work we will explore the behaviour and implications of utilizing two types of lower fidelity proxies in conjunction with model based perfor- mance predictor. The first type of lower fidelity proxies being zero cost (ZC) proxies used as additional input features for the model besides of standard architecture's encod- ing. The second type being learning curve extrapolation used for generating labels of the model based predictor's training dataset hence compensating for its otherwise very long initialization time. 1
|
|
|
Collision Avoidance in Computer Games
Lakatoš, Peter ; Gemrot, Jakub (advisor) ; Pilát, Martin (referee)
Collision avoidance for autonomous agents has been a widely researched topic for the past couple of decades. Modern solutions act as purely reactive techniques that create various problems, such as agents being stuck in various scenarios. The aim of this thesis was to explore a new way of solving collision avoidance for humanoid agents using genetic algorithms to search local space multiple steps ahead of the current simulation state. The application is capable of running multiple predefined test scenarios and logging the results of each run. The application provides two possible ways of seeing the results, either visually observing the scenario run or plotting the results logged by the application. The overall design of the application is general enough to allow simple modification to existing scenarios or creation of new ones. It is also possible to modify an existing genetic algorithm with new operators with minimal effort. The results show that even though various configurations of the implemented genetic algorithm perform similarly, there are some outstanding winners that might bring an alternative possibility to the already existing collision avoidance methods.
|
|
|
Multi-agent trading environment for training robust reinforcement learning agents
Mikuláš, Pavel ; Pilát, Martin (advisor) ; Neruda, Roman (referee)
This thesis presents a comprehensive study of the application of reinforcement learning to algorithmic trading. The main focus of this thesis is on the generalization properties of various reinforcement learning algorithms, both from the data perspective and the applicability of the trained agents to real algorithmic trading. To that end, we develop a training environment taking into account various real-world factors influencing the performance of algorithmic trading strategies. We also experiment with the recurrent replay buffer extension of the DQN algorithm, known as R2D2, being, to the best of our knowledge, the first to employ this algorithm for the task of algorithmic trading. Each algorithm is evaluated against traditional algorithmic trading strategies, including the buy-and-hold strategy, to demonstrate the superior performance of the reinforcement learning strategies. On top of that we also provide a study on how the amount of training data and transaction costs influence the generalization of the algorithms to unseen market conditions. We show how transaction costs significantly increase the task complexity and that the R2D2 algorithm overperforms the commonly used baselines, as well as other state-of-the-art reinforcement learning algorithms in this task. 1
|
|
|
Evolution strategies for policy optimization in transformers
Lorenc, Matyáš ; Neruda, Roman (advisor) ; Pilát, Martin (referee)
We explore the capability of evolution strategies to train a transformer architecture in the reinforcement learning setting. We perform experiments using OpenAI's highly parallelizable evolution strategy and its derivatives utilizing novelty and quality-diversity searches to train Decision Transformer in Humanoid locomotion environment, testing the ability of these black-box optimization techniques to train even such relatively large (com- pared to the previously tested in the literature) and complicated (using a self-attention in addition to fully connected layers) models. The tested algorithms proved to be, in gen- eral, capable of achieving strong results and managed to obtain high-performing agents both from scratch (randomly initialized model) and from a pretrained model. 1
|
|
|
Diplomacy-Based Strategy Game
Valach, Miroslav ; Pilát, Martin (advisor) ; Guba, Peter (referee)
Strategy games are known for allowing players to choose from a vast array of different strategies that can be employed to achieve victory. The major- ity of these games revolve around the standardized pillars of 4X (Explore, Expand, Exploit, Exterminate) games such as Civilization or Stellaris. How- ever, these pillars often encourage conquest or aggressive means to achieve victory, thereby rendering a peaceful approach as a rarely viable strategy to pursue. In this project, our aim was to create a strategy game that focuses on player interaction through diplomacy. The main goal is to provide players with free- dom similar to Diplomacy, where players can and have to utilize multilateral politics in order to achieve victory. As a proof of concept, we have successfully developed a game prototype using Unreal Engine. The prototype showcases a diplomacy-based game with multilateral diplomacy at its core. The gameplay demonstrates the viability of diplomacy as the primary strategy for achieving victory in video games.
|
|
|
Creating Adversarial Examples in Machine Learning
Červíčková, Věra ; Pilát, Martin (advisor)
This thesis examines adversarial examples in machine learning, specifically in the im- age classification domain. State-of-the-art deep learning models are able to recognize patterns better than humans. However, we can significantly reduce the model's accu- racy by adding imperceptible, yet intentionally harmful noise. This work investigates various methods of creating adversarial images as well as techniques that aim to defend deep learning models against these malicious inputs. We choose one of the contemporary defenses and design an attack that utilizes evolutionary algorithms to deceive it. Our experiments show an interesting difference between adversarial images created by evolu- tion and images created with the knowledge of gradients. Last but not least, we test the transferability of our created samples between various deep learning models. 1
|
|
|
Football Player Performance Prediction
Kellich, Adam ; Pilát, Martin (advisor) ; Pešková, Klára (referee)
This thesis focuses on the development of tools to improve the experience of play- ing the online fantasy football game Sorare. In Sorare, players buy collectible cards that represent real footballers and compete against other players, with success depending on the actual performance of the footballers in real matches. Our work aims to address two main problems that players face in the game. First, we attempt to accurately predict a soccer player's score in an upcoming match based on data available before the match. Second, we seek to identify players undervalued by the market, which represent potential investment opportunities. We describe the whole process of machine learning from data acquisition, data processing, designing appropriate algorithms, training models to eva- luation. In both cases, the proposed algorithms have demonstrated usefulness compared to simple average-based predictions. 1
|
guest ::
RSS feed.