Search

U ovom radu opisan je sustav za konstruiranje i upravljanje portfeljom dionica. Konstruiranje diversificiranog portfelja dionica je napravljeno korištenjem konvolucijskog autoenkodera i Leiden-ovog algoritma za otkrivanje zajednica. Upravljanje konstruiranim portfeljom dionica napravljeno pomoću metoda iz dubokog potpornog učenja. Arhitekutra agenta koji je donosio odluke je bila izmjenjena arhitektura Transformera. Opisani sustav je ispitan na vremenskim nizovima cijena dionica koje pripadaju S&P 500 indexu. This paper describes a system for constructing and managing a stock portfolio. The construction of a diversified stock portfolio was done using a convolutional autoencoder and Leiden’s community detection algorithm. Stock portfolio managment was done using methods from deep reinforcement learning. The architecture of agent in deep reinforcement learning was modified architecture of the Transformer. The described system has been tested on stock price time series belonging to the S&P 500 index

U ovom diplomskom radu mjeri se uspješnost primjene algoritama dubokog potpornog učenja za pronalazak rješenja problema Inverzne kinematike zglobova manipulatora. Eksperimentalni dio se provodi na robotu Baxteru koji ima sedam stupnjeva slobode. Inverzna kinematika je problem koji se sastoji od pronalaženja orijentacije i pozicije pojedinog zgloba kako bi manipulator došao u željenu točku. Analitičko rješenje ovog problema ne postoji stoga su se algoritmi dubokog potpornog učenja pokazali kao dobra metoda za pronalazak adekvatnog rješenja. U radu su opisane opće karakteristike robota, metoda učenja, primijenjeni algoritam i na kraju dobiveni eksperimentalni rezultati. This thesis measures the success of using deep reinforcement learning algorithms for finding solutions to the Inverse kinematics problem. The experimental part is performed on robot Baxter that has seven degrees of freedom. Inverse kinematics is a problem that consists of finding the orientation and position of an individual joint in order for the manipulator to reach the desired point.There is no analytical solution to this problem, so deep reinforcement learning algorithms have proven to be a good method for finding an adequate solution. The paper describes the general characteristics of robots, learning methods, applied algorithm and finally obtained experimental results.

U 2017. godini je DeepMind izdao rad u kojem je opisao glavnu ideju kako ostvariti šahovski stroj temeljen na dubokom podržanom učenju. Ovaj rad se bavi implementiranjem opisanog algoritma na jedinstven način. Algoritam se temelji na dubokom podržanom učenju, odnosno na neuronskim mrežama i algoritmima podržanog učenja. U izradi se koriste rezidualne neuronske mreže koje omogućavaju izradu vrlo dubokih mreža, tj. mreža s puno slojeva. Korišteni algoritam podržanog učenja je obilazak stabla tehnikom Monte Carlo. Implementacijom neuronskih mreža i Monte Carlo algoritma se gradi model šahovskog stroja te se on uči koji su potezi dobri na temelju nagrada i kazni koje dobije u partijama koje igra sam sa sobom. Izgrađeni stroj ima još dosta mjesta za nadogradnje. In 2017, DeepMind published a paper describing the main idea of how to realize a chess machine based on deep reinforcement learning. This paper deals with the implementation of the described algorithm in a unique way. The algorithm is based on deep reinforcement learning, i.e. neural networks and reinforcement learning algorithms. Residual neural networks are used in designing the model, which enable very deep networks, i.e. networks with many layers. The supported learning algorithm used is a tree tour using the Monte Carlo technique. By implementing neural networks and the Monte Carlo algorithm, a model of a chess machine is built and one learns which moves are good based on the reward and punishment received in the games they play with. The built machine still has plenty of room to upgrade.

Postavljena je simulacija u programskom sustavu Unity. Uporabljene su tri različite scene za učenje. U sceni se nalazi dron koji se može gibati u dvije dimenzije, s fiksiranom visinom. Naučen je model arhitekture Pix2Pix za mono-okularnu predikciju dubine.. Korišteni su podaci generirani u okviru simulacije. Algoritam izbjegavanja prepreka je Q-učenje. Simulacija daje algoritmu RGB sliku prostora ispred drona. Predviđa se dubina i zajedno sa željenim smjerom daje se kao ulaz u duboku Q-mrežu (DQN). Model se sastoji od potpuno povezanih i konvolucijskih slojeva. Ostvarena je i varijanta s LSTM slojem. Radi usporedbe su korištene i neke heuristike za izbjegavanje prepreka. Najbolji rezultat je postignut kombinacijom predviđene dubine i varijantom algoritma polja potencijala. A simulation is built using Unity. There are three different testing scenes. A drone is placed in the scene. It can only move in two dimensions. The height is fixed. Pix2Pix architecture is used for single camera depth prediction. Training dataset is generated from the simulation. Algorithm used for obstacle evasion is Q-learning. Simulation provides algorithm with RGB image of the area in front of the drone. Predicted depth with target angle is passed as input to a Deep Q-Network (DQN). Model is composed of dense and convolution layers. There is also a DQN version which uses a LSTM layer. Several obstacle avoidance heuristics are developed to better compare results of the DQN. The best result is achieved by using predicted depth with potential field algorithm.

Cilj ovog rada je napraviti simulaciju lanca opskrbe pivomi računalne agente koji minimiziraju vlastite troškove unutar tog lanca. Opisana je problematika lanca opskrbe, duboko podržanoučenjei korištenialgoritam.Objašnjena je i implementacija programskog rješenja za zadani problem. Nakon provedenog učenja agenata, opisani su dobiveni rezultatii njihova prednost nad drugim postojećim rješenjima. The goal of this paper is to make a simulation of a beer supply chain and a computer agentsthat minimize its own cost within that chain. Supply chain issues, deepreinforcementlearning and theusedalgorithmare described. The implementation of the software solution for the given problem is also explained. After the proven learning of the agents, the obtained results and their advantage over other existing solutions are described.

Inverzna kinematika manipulatora učenja iznimno je kompliciran problem. Cilj ovog rada je bio riješiti taj problem uz pomoć dubokog potpornog učenja. Napravljen je pregled područja dubokog potpornog učenja, iznesene su teoretske osnove, ključni pojmovi dubokog potpornog učenja te algoritmi gradijenta politike. Riješen je problem dohvaćanja objekta u prostoru korištenjem robotske ruke Jaco, prvo u prostoru bez prepreka, a zatim i s preprekama. Inverse kinematics of robotic manipulators is a very complicated problem. The aim of this thesis was to solve that problem using deep reinforcement learning. An overview of the area of deep reinforcement learning was made, the theoretical foundations, key concepts of deep supportive learning and policy gradient algorithms were presented. Problem of reaching target in space using robotic manipulator arm Jaco was solved, first in space without obstacles and then in space with obstacles.

Ovaj rad prikazuje rezultate evaluacije primjene algoritama podržanog učenja na problem učenja lokomocije jednostavnog simuliranog tijela. Osim prikaza i analize rezultata, ovaj rad sadrži i detaljan uvod u polje podržanog učenja. Uvode se osnove klasičnog podržanog učenja i matematičke podloge istoga. Zatim se pruža pregled osnovnih algoritama klasičnog podržanog učenja. Uvode se i osnove dubokog učenja kako bi čitatelj razvio intuiciju o primjeni takvih modela u podržanom učenju. Također, detaljno se opisuje i svaki korišteni algoritam. This paper presents the results of the evaluation of the application of reinforcement learning algorithms to the locomotion learning problem of a simple simulated body. In addition to the presentation and analysis of results, this paper contains a detailed introduction to the field of reinforcement learning. The basics of classic reinforcement learning and its mathematical basis are introduced. An overview of the basic algorithms of classical supported learning is then provided. The basics of deep learning are also introduced to help the reader develop an intuition about the application of such models in supported learning. Also, each used algorithm is described in detail.

U ovom radu objašnjena su teorijska načela potrebna za shvaćanje osnovne problematike dubokog podržanog učenja i izučavani su različiti algoritmi kojima je moguće riješiti problem dubokog podržanog učenja s primjenom na video igre. Algoritmi koji su obrađeni u okviru rada su: duboka Q mreža, dvostruka duboka Q mreža, suparnička dvostruka duboka Q mreža te asinkroni prednosni akter-kritičar. U okviru rada su implementirani svi obrađeni algoritmi, provedena su mjerenja nad njima na temelju igre Atari Breakout i uspoređeni su rezultati mjerenja. Rezultati mjerenja pokazuju da je asinkroni prednosni akter-kritičar algoritam koji nudi najbolje performanse s obzirom na vremenske i sklopovske resurse koji su potrebni pri rješavanju problema dubokog podržanog učenja. This thesis lays out theoretical principles that are needed in order to understand the problem of deep reinforcement learning and studies various algorithms that can solve the problem of deep reinforcement learning with application to video games. Algorithms that are covered in this thesis are: Deep Q Network, Double Deep Q Network, Dueling Double Deep Q Network and asynchronous advantage actor-critic. Aforementioned algorithms were implemented within this thesis, measurements were taken with respect to the Atari Breakout video game and the measurement results were compared. Results show that asynchronous advantage actor-critic algorithm shows the best performance with respect to time and hardware resources that are needed for solving deep reinforcement learning task.