|
|
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
|
|
|
Detekce skupin v davech pomocí časoprostorových dat
Říha, David ; Hartman, David (advisor) ; Neruda, Roman (referee)
This thesis addresses the challenge of social group detection in crowds, presenting an algorithm informed by sociological insights into common group formations among pedestrians. Our proposed algorithm demonstrates comparable performance to existing solutions - Time-sequence DBSCAN and Agglomerative Hierarchical Clustering with Hausdorff Distance, using the DIAMOR dataset for testing and comparison. Additionally, we introduce a validator tool potentially capable of refining results from existing algorithms based on a group shape criterion, leading to improved accuracy in identifying groups. Keywords: groups detection; clustering; group shape analysis; pedestrian behavior;
|
|
|
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
|
|
|
Epidemiologické modely s agenty
Neruda, Roman
Tento příspěvek je jemným úvodem do problematiky agentních modelů a jejich aplikací v epidemiologickém modelování. Představíme agentní modely jednak z hlediska informatiky, jednak jako nástroj modelování v jiných vědních disciplínách. V příkladové studii ukážeme model s agenty a sociální sítí jejich kontaktů, který slouží pro simulaci vývoje epidemie a vlivu protiepidemických opatření.
|
|
|
Predicting Production Times Using Machine Learning
Novotný, Tomáš ; Pilát, Martin (advisor) ; Neruda, Roman (referee)
With the development of automated task planning in industry, the requirements for a correct estimation of the parameters of individual operations, especially their lead time, are increasing. This thesis is discussing various methods of estimating the lead time for new tasks automatically from previously executed tasks. The first part of the thesis focuses on standard regression algorithms and their modifications according to the suitability for this problem. The second, main part of the theses, focuses on solutions using deep neural networks, which are able to process unstructured data, such as textual descriptions of operations. The final results show that deep learning achieves a good level of prediction, especially for new types of operations. Its practical use can therefore be recommended as an estimate for planning new products, especially in highly dynamic environments.
|
|
|
Image Reassembling Algorithms
Yamalutdinova, Yuliya ; Pilát, Martin (advisor) ; Neruda, Roman (referee)
Jigsaw puzzle is a well-known puzzle game that has been around for centuries. How- ever, in addition to entertainment purposes, an ability to reassemble images from pieces has practical applications and can be useful, for example, in restoring torn or cut docu- ments or broken objects in archaeology. Most of the proposed solutions reconstruct the images divided into square pieces. In this work, we propose our solutions for the new types of puzzles with more interesting shapes of pieces, such as rectangles of equal and different sizes, and triangles. The accuracy of our solvers is at the same level as that of the solutions for reconstruction of images from square pieces. Moreover, our solvers, unlike the others, have been tested on images with text as well as on color and black and white photographs. 1
|
|
|
Utilization of brain connectivity in classification and regression tasks in brain data
Řežábková, Jana ; Hartman, David (advisor) ; Neruda, Roman (referee)
This thesis investigates how incorporating progressive amounts of struc- tural information into machine learning models affects their accuracy in dis- criminating schizophrenia from functional connectivity matrices obtained by resting state functional magnetic resonance. Three structural settings were explored-no structure via traditional machine learning models, modeling nodes through proposed feed forward based architecture that allows com- bining node neighborhoods individually for each node, and modeling both nodes and edges using graph neural networks. Although the results on the available 190 subjects dataset did not reveal the best strategy, two findings were identified (a) the superiority of sparsifiying matrices by taking top k neighborhoods over keeping top n% values and (b) the benefit of node cor- respondence across samples. All experiments were evaluated using a proper validation strategy-nested cross validation-a piece that was largely missing in reviewed literature.
|
|
|
Deep Learning for Symbolic Regression
Vastl, Martin ; Pilát, Martin (advisor) ; Neruda, Roman (referee)
Symbolic regression is a task of finding mathematical equation based on the observed data. Historically, genetic programming was the main tool to tackle the symbolic regres- sion, however, recently, new neural network based approaches emerged. In this work, we propose transformer based approach which predicts the expression as a whole without the need of finding the expression coefficients in post-processing step. We also use a local gradient search to further improve the expression coefficients. We compare our so- lution to previous approaches on several benchmarks and demonstrate, that our solution is comparable in terms of performance while outperforming them in terms of speed of the prediction in the average case. 1
|
|
|
Quasi-parallel optimizing of simulated systems by means of genetic algorithms
Konopa, Michal ; Kindler, Evžen (advisor) ; Neruda, Roman (referee)
This work goals are SIMULA classes to model experts sessions so that each expert has his own start ideas the optimal system's parameters. All the experts are simulating the system (each expert with his own parameters), but during the simulation they are mutually exchanging information about behavior of theirs models and - in accordance with this information - they are learning, changing their own system parameters. The learning process is performed by means of genetic algorithms. The resulting optimizer is tested on concrete examples, both from the mathematical theory and real practice (e.g. optimizing of a given project). This work reassumes the dissertation of RNDR. Jiří Weinberger. CSc., who made similar type of optimizer in the 80es, when nothing concrete was known on genetic algorithms and so the experts learning scheme was modeled by techniques, which had great success in solution of real problems afterwards; but nowadays it is worth to replace them by genetic algorithms, or at least to compare both methods. Genetic algorithms are successfully used in systems optimization, where the model run is algorithmic-managed, but never in the way of directly changing the models during their running times. Parallel run of multiple simulating models on the computer equipped with only one processor requires...
|
|
|
Agent optimization by means of genetic programming
Šmíd, Jakub ; Neruda, Roman (advisor) ; Kazík, Ondřej (referee)
This thesis deals with a problem of choosing the most suitable agent for a new data mining task not yet seen by the agents. The metric is proposed on the data mining tasks space, and based on this metric similar tasks are identified. This set is advanced as an input to a program evolved by means of genetic programming. The program estimates agents performance on the new task from both the time and error point of view. A JADE agent is implemented which provides an interface allowing other agents to obtain estimation results in real time.
|
guest ::
