56 research outputs found
Echo state networks and their application in time series prediction
No full textRecurrent neural networks (RNN) enable to model dynamical sys- tems with variable input length. Their disadvantage is in inherently difficult trai- ning which means adjusting weights of connections between neurons connected in the network. Echo state networks (ESN) are a special type of RNN which are by contrast trainable rather simply. They include a reservoir of neurons whose state reflect the history of all signals in the network and that is why this type of network is suitable for simulation and prediction of time series. To maximize the computational power of ESN, very precise adjusting and experimenting are required. Because of that, we have created a tool for building and testing such networks. We have implemented a time series forecasting task for the purpose of examination of our tool. We have focused on stock price prediction, which repre- sents an uncertain and complicated area for achieving precise results in. We have compared our tool to other tools and it was comparably successful
Vytváření herních strategií pro hru PuppetWars pomocí neuroevoluce
No full textIn recent years the gaming industry has been on increase. In order to maintain competitiveness gaming companies are required to develop more and more compelling computer games what implies the presence of the very responsive artificial intelligence controlling the game elements, on which our work focuses. We have implemented a simple 2D programming game where we have experimented with the artificial intelligence in it trying to create a strategy beeing able to compete with human. We have explored several variations of learning through the evolutionary strategy applied to neural networks and we have created game characters worthy of being an equal opponent to the game user.V posledných rokoch nastal v hernom priemysle rozmach. Pre udržanie konkurencieschopnosti sú herné spoločnosti nútené vyvíjať stále viac príťažli- vé počítačové hry, čo implikuje i prítomnosť čo najvernejšej umelej inteligen- cie ovládajúcej herné prvky, na čo sa naša práca zameriava. Implementovali sme jednoduchú 2D programovaciu hru, na ktorej sme predviedli sadu poku- sov učiac umelú inteligenciu v nej, snažiac sa vytvoriť stratégie konkurujúce tým ľudským. Preskúmali sme niekoľko variácií učenia pomocou evolučnej stratégie aplikovanej na neurónové siete a vytvorili sme herné postavičky hodné bytia rovnocenným protivníkom užívateľom hry.Katedra teoretické informatiky a matematické logikyDepartment of Theoretical Computer Science and Mathematical LogicFaculty of Mathematics and PhysicsMatematicko-fyzikální fakult
Strojové učení na malých datových množinách s velkým počtem atributů
No full textMachine learning models are difficult to employ in biology-related research. On the one hand, the availability of features increases as we can obtain gene expressions and other omics information. On the other hand, the number of available observations is still low due to the high costs associated with obtaining the data for a single subject. In this work we, therefore, focus on the set of problems where the number of observations is smaller than the number of features. We analyse different combinations of feature selection and classification models and we study which combinations work the best. To assess these model combinations, we introduce two simulation studies and several real-world datasets. We conclude that most classification models benefit from feature pre-selection using feature selection models. Also, we define model-based thresholds for the number of observations above which we observe increased feature selection stability and quality. Finally, we identify a relation between feature selection False Discovery Rate and stability expressed in terms of the Jaccard index. 1Použitie metód strojového učenia v oblasti biológie je často zložité. Na jednej strane sa v tejto oblasti zvyšuje počet premenných, ktoré je možné zaznamenať. Na strane druhej je však meranie pre každú entitu nákladné, a tak je počet pozorovaní často nízky. V tejto práci sa preto venujeme datasetom, ktoré majú nízky počet pozorovaní, ale veľký počet premenných. Zameriavame sa na rozličné kombinácie metód výberu premenných a klasifikačných metód a pokúšame sa odpovedať na otázku, ktoré kombinácie fungujú najlepšie. Pre lepšie pochopenie súvislostí medzi metódami pracujeme s dvoma simu- lačnými štúdiami a niekoľkými reálnymi datasetmi. Výsledky práce naznačujú, že väčšina klasifikačných metód dosahuje lepšie výsledky, ak pracujú s predvybranými premennými a nie so všetkými premennými dostupnými v dátach. V práci zároveň definujeme hranice pre počet pozorovaní datasetu, od ktorých dosahujú metódy výberu premenných vyššiu kvalitu a stabilitu. V závere práce popisujeme identifikovaný vzťah medzi mierou stabil- ity výberu premenných (tzv. Jaccardov index) a mierou kvality výberu premenných (tzv. False Discovery Rate). 1Department of Software and Computer Science EducationKatedra softwaru a výuky informatikyMatematicko-fyzikální fakultaFaculty of Mathematics and Physic
Creating the Game Strategies for PuppetWars Using Neuroevolution
No full textIn recent years the gaming industry has been on increase. In order to maintain competitiveness gaming companies are required to develop more and more compelling computer games what implies the presence of the very responsive artificial intelligence controlling the game elements, on which our work focuses. We have implemented a simple 2D programming game where we have experimented with the artificial intelligence in it trying to create a strategy beeing able to compete with human. We have explored several variations of learning through the evolutionary strategy applied to neural networks and we have created game characters worthy of being an equal opponent to the game user
Optimization of a circulating multi-car elevator system
No full textCirculating multi-car elevator is a system holding multiple cars in two shafts, where cars move upwards in one shaft and downwards in the other shaft. This system is similar to the paternoster, but cars have to stop on floors and open doors to load and unload passengers. Besides many technical challenges, this sys- tem brings algorithmic problems regarding efficient control of all cars. This thesis studies an off-line optimization problem, where the most efficient elevator system is searched for a fixed set of passengers. For this purpose, we created a computer program, implementing a genetic algorithm for searching for the most efficient elevator control and a discrete event simulation for evaluation of the efficiency of the control. The program provides a graphical user interface for input of parame- ters, generating passengers and displaying the results.
Machine learning on small datasets with large number of features
No full textMachine learning models are difficult to employ in biology-related research. On the one hand, the availability of features increases as we can obtain gene expressions and other omics information. On the other hand, the number of available observations is still low due to the high costs associated with obtaining the data for a single subject. In this work we, therefore, focus on the set of problems where the number of observations is smaller than the number of features. We analyse different combinations of feature selection and classification models and we study which combinations work the best. To assess these model combinations, we introduce two simulation studies and several real-world datasets. We conclude that most classification models benefit from feature pre-selection using feature selection models. Also, we define model-based thresholds for the number of observations above which we observe increased feature selection stability and quality. Finally, we identify a relation between feature selection False Discovery Rate and stability expressed in terms of the Jaccard index.
Echo state sítě a jejich využití na předpovídání časových řad
No full textRecurrent neural networks (RNN) enable to model dynamical sys- tems with variable input length. Their disadvantage is in inherently difficult trai- ning which means adjusting weights of connections between neurons connected in the network. Echo state networks (ESN) are a special type of RNN which are by contrast trainable rather simply. They include a reservoir of neurons whose state reflect the history of all signals in the network and that is why this type of network is suitable for simulation and prediction of time series. To maximize the computational power of ESN, very precise adjusting and experimenting are required. Because of that, we have created a tool for building and testing such networks. We have implemented a time series forecasting task for the purpose of examination of our tool. We have focused on stock price prediction, which repre- sents an uncertain and complicated area for achieving precise results in. We have compared our tool to other tools and it was comparably successful. 1Rekurentné neurónové siete (RNN) umožňujú modelovať dynamické systémy s premenlivou dĺžkou vstupu. Ich nevýhoda je v náročnom učení, teda ťažkom nastavovaní váh neurónov, ktoré sú v sieti spojené. Echo state siete (ESN) sú zvláštnym typom RNN, ktoré sa naopak dajú učiť relatívne jednoducho. Majú rezervoár neurónov, ktorých stav odráža históriu všetkých signálov v sieti, a preto je tento typ sietí vhodný na simuláciu a predikciu časových postupností. Dosia- hnutie maximálneho výkonu ESN vyžaduje veľmi precízne nastavenie a experi- mentovanie. Preto sme v tejto práci vytvorili nástroj pre konštruovanie a testova- nie takýchto sietí. Pre účely vyskúšania vytvoreného nástroja sme implementovali úlohu predpovedania vývoja časových radov. Konkrétne sme sa zamerali na pred- povedanie vývoja cien akcií, čo predstavuje veľmi neistú a pre presnú predpoveď veľmi náročnú oblasť. Výsledky našich experimentov sme porovnali s výsledkami iných nástrojov, a zistili sme, že náš nástroj dáva porovnateľné výsledky. 1Katedra softwaru a výuky informatikyDepartment of Software and Computer Science EducationMatematicko-fyzikální fakultaFaculty of Mathematics and Physic
Echo state siete a ich využitie na predpovedanie časových radov
No full textRecurrent neural networks (RNN) enable to model dynamical sys- tems with variable input length. Their disadvantage is in inherently difficult trai- ning which means adjusting weights of connections between neurons connected in the network. Echo state networks (ESN) are a special type of RNN which are by contrast trainable rather simply. They include a reservoir of neurons whose state reflect the history of all signals in the network and that is why this type of network is suitable for simulation and prediction of time series. To maximize the computational power of ESN, very precise adjusting and experimenting are required. Because of that, we have created a tool for building and testing such networks. We have implemented a time series forecasting task for the purpose of examination of our tool. We have focused on stock price prediction, which repre- sents an uncertain and complicated area for achieving precise results in. We have compared our tool to other tools and it was comparably successful.Rekurentné neurónové siete (RNN) umožňujú modelovať dynamické systémy s premenlivou dĺžkou vstupu. Ich nevýhoda je v náročnom učení, teda ťažkom nastavovaní váh neurónov, ktoré sú v sieti spojené. Echo state siete (ESN) sú zvláštnym typom RNN, ktoré sa naopak dajú učiť relatívne jednoducho. Majú rezervoár neurónov, ktorých stav odráža históriu všetkých signálov v sieti, a preto je tento typ sietí vhodný na simuláciu a predikciu časových postupností. Dosia- hnutie maximálneho výkonu ESN vyžaduje veľmi precízne nastavenie a experi- mentovanie. Preto sme v tejto práci vytvorili nástroj pre konštruovanie a testova- nie takýchto sietí. Pre účely vyskúšania vytvoreného nástroja sme implementovali úlohu predpovedania vývoja časových radov. Konkrétne sme sa zamerali na pred- povedanie vývoja cien akcií, čo predstavuje veľmi neistú a pre presnú predpoveď veľmi náročnú oblasť. Výsledky našich experimentov sme porovnali s výsledkami iných nástrojov, a zistili sme, že náš nástroj dáva porovnateľné výsledky.Katedra softwaru a výuky informatikyDepartment of Software and Computer Science EducationFaculty of Mathematics and PhysicsMatematicko-fyzikální fakult
Controlled Music Generation with Deep Learning
No full textGeneration of musical compositions is one of the hardest tasks for artificial intelligence where most of the current approaches struggle with long term coherence of the generated compositions. This work aims to demonstrate how deep learning models for generating music can be externally controlled to produce compositions with long term coherence, polyphony, and multiple instruments. We work with classical music ranging from compositions for piano through string quartet and up to symphonic orchestral compositions. To control the generation process, we take inspiration from the abstract notion of musical form: normally a high-level description of how similar and dissimilar passages are arranged throughout a composition, we use it as a recipe for generating a coherent composition. To this end, we (1) design a sufficiently general music similarity pseudometric from existing methods, (2) extract musical form from the training data by applying a clustering algorithm over the similarity values, (3) train three models that generate similar and locally coherent dissimilar musical fragments, and (4) design a way how to use the musical forms during the generation process to orchestrate the inference of the three models to generate whole compositions from musical fragments. We show what is the performance of the..
Protein secondary structure prediction using deep neural networks
No full textDetermination of protein structure in space is a crucial part of protein function analysis. But structure determination is an expensive and time consuming pro- cess, therefore structure prediction model raised on popularity. The most notable subproblem of protein structure prediction is prediction of local conformation of the adjacent amino acids, ie. secondary structure. This thesis studies usage of deep neural networks for protein secondary structure prediction. We implemented pre- diction model and different modifications are evaluated. Especially compassion of LSTM and GRU memory cells was done. Furthermore, two new preprocessing me- thods are evaluated. Fast PSSM calculation method was proposed and prediction of tertiary structure was used as input for prediction model. Last part of this thesis examine application of filtering methods for models predicting secondary structure with eight classes.
- …
