1 |
Self-adjusting reinforcement learningDer, Ralf, Herrmann, Michael 10 December 2018 (has links)
We present a variant of the Q-learning algorithm with automatic control of the exploration rate by a competition scheme. The theoretical approach is accompanied by systematic simulations of a chaos control
task. Finally, we give interpretations of the algorithm in the context of computational ecology and neural networks.
|
2 |
Improving Time Efficiency of Feedforward Neural Network LearningBatbayar, Batsukh, S3099885@student.rmit.edu.au January 2009 (has links)
Feedforward neural networks have been widely studied and used in many applications in science and engineering. The training of this type of networks is mainly undertaken using the well-known backpropagation based learning algorithms. One major problem with this type of algorithms is the slow training convergence speed, which hinders their applications. In order to improve the training convergence speed of this type of algorithms, many researchers have developed different improvements and enhancements. However, the slow convergence problem has not been fully addressed. This thesis makes several contributions by proposing new backpropagation learning algorithms based on the terminal attractor concept to improve the existing backpropagation learning algorithms such as the gradient descent and Levenberg-Marquardt algorithms. These new algorithms enable fast convergence both at a distance from and in a close range of the ideal weights. In particular, a new fast convergence mechanism is proposed which is based on the fast terminal attractor concept. Comprehensive simulation studies are undertaken to demonstrate the effectiveness of the proposed backpropagataion algorithms with terminal attractors. Finally, three practical application cases of time series forecasting, character recognition and image interpolation are chosen to show the practicality and usefulness of the proposed learning algorithms with comprehensive comparative studies with existing algorithms.
|
3 |
Testing Safety Critical Avionics Software Using LBTestStenlund, Sebastian January 2016 (has links)
A case study for the tool LBTest illustrating benets and limitations of the tool along the terms of usability, results and costs. The study shows the use of learning based testing on a safety critical application in the avionics industry. While requiring the user to have the oretical knowledge of the tools inner workings, the process of using the tool has benefits in terms of requirement analysis and the possibility of finding design and implementation errors in both the early and late stages of development
|
4 |
Faster Adaptive Network Based Fuzzy Inference SystemWeeraprajak, Issarest January 2007 (has links)
It has been shown by Roger Jang in his paper titled "Adaptive-network-based fuzzy inference systems" that the Adaptive Network based Fuzzy Inference System can model nonlinear functions, identify nonlinear components in a control system, and predict a chaotic time series. The system use hybrid-learning procedure which employs the back-propagation-type gradient descent algorithm and the least squares estimator to estimate parameters of the model. However the learning procedure has several shortcomings due to the fact that * There is a harmful and unforeseeable influence of the size of the partial derivative on the weight step in the back-propagation-type gradient descent algorithm. *In some cases the matrices in the least square estimator can be ill-conditioned. *Several estimators are known which dominate, or outperform, the least square estimator. Therefore this thesis develops a new system that overcomes the above problems, which is called the "Faster Adaptive Network Fuzzy Inference System" (FANFIS). The new system in this thesis is shown to significantly out perform the existing method in predicting a chaotic time series , modelling a three-input nonlinear function and identifying dynamical systems. We also use FANFIS to predict five major stock closing prices in New Zealand namely Air New Zealand "A" Ltd., Brierley Investments Ltd., Carter Holt Harvey Ltd., Lion Nathan Ltd. and Telecom Corporation of New Zealand Ltd. The result shows that the new system out performed other competing models and by using simple trading strategy, profitable forecasting is possible.
|
5 |
Neuromorphic systems for legged robot controlMonteiro, Hugo Alexandre Pereira January 2013 (has links)
Locomotion automation is a very challenging and complex problem to solve. Besides the obvious navigation problems, there are also problems regarding the environment in which navigation has to be performed. Terrains with obstacles such as rocks, steps or high inclinations, among others, pose serious difficulties to normal wheeled vehicles. The flexibility of legged locomotion is ideal for these types of terrains but this alternate form of locomotion brings with it its own challenges to be solved, caused by the high number of degrees of freedom inherent to it. This problem is usually computationally intensive, so an alternative, using simple and hardware amenable bio-inspired systems, was studied. The goal of this thesis was to investigate if using a biologically inspired learning algorithm, integrated in a fully biologically inspired system, can improve its performance on irregular terrain by adapting its gait to deal with obstacles in its path. At first, two different versions of a learning algorithm based on unsupervised reinforcement learning were developed and evaluated. These systems worked by correlating different events and using them to adjust the behaviour of the system so that it predicts difficult situations and adapts to them beforehand. The difference between these versions was the implementation of a mechanism that allowed for some correlations to be forgotten and suppressed by stronger ones. Secondly, a depth from motion system was tested with unsatisfactory results. The source of the problems are analysed and discussed. An alternative system based on stereo vision was implemented, together with an obstacle detection system based on neuron and synaptic models. It is shown that this system is able to detect obstacles in the path of the robot. After the individual systems were completed, they were integrated together and the system performance was evaluated in a series of 3D simulations using various scenarios. These simulations allowed to conclude that both learning systems were able to adapt to simple scenarios but only the one capable of forgetting past correlations was able to adjust correctly in the more complex experiments.
|
6 |
Apprentissage non-supervisé dans les modèles linéaires gaussiens. Application à la biométrie dynamique de l’iris / Unsupervised Learning in linear Gaussian models. Application to the dynamic iris biometricsNemesin, Valérian 30 September 2014 (has links)
La famille de modèles dite des filtres de Kalman permet d'estimer les états d'un système dynamique à partir d'une série de mesures incomplètes ou bruitées. Malgré leur relative simplicité de modélisation, ces filtres sont utilisés dans un large spectre scientifique dont le radar, la vision, et les communications. Ce succès repose, pour l'essentiel, sur l'existence d'algorithmes de filtrage et de lissage exacts et rapides, \ie linéaires au nombre d'observations, qui minimisent l'erreur quadratique moyenne. Dans cette thèse, nous nous sommes intéressés au filtre de Kalman couple. Celui-ci intègre, par rapport au modèle original, de nouvelles possibilités d'interactions entre états cachés et observations, tout en conservant des algorithmes exacts et rapides dans le cas linéaire et gaussien. Nous étudions plus particulièrement le problème de l'estimation non supervisée et robuste des paramètres d'un filtre de Kalman couple à partir d'observations en nombre limité. Le manuscrit décrit ainsi plusieurs algorithmes d'apprentissage par estimation du maximum de vraisemblance selon le principe EM (\textit{Expectation-Maximization}). Ces algorithmes originaux permettent d'intégrer des contraintes a priori sur les paramètres du système étudié, comme expressions de connaissances partielles sur la physique de l'application ou sur le capteur. Ces systèmes contraints réduisent l'ambiguïté liée au problème d'identifiabilité du filtre de Kalman couple lors de l'estimation des paramètres. Ils permettent également de limiter le nombre de maxima locaux de la fonction de vraisemblance en réduisant la dimension de l'espace de recherche, et ainsi évitent parfois le piégeage de l'algorithme EM. Il est important de noter que l'ensemble des algorithmes proposés dans ce manuscrit s'applique directement au filtre de Kalman original, comme cas particulier du filtre de Kalman couple. Tous les algorithmes sont rendus robustes par la propagation systématique de racines-carrés des matrices de covariance au lieu des matrices de covariance elles-mêmes, permettant ainsi d'éviter les difficultés numériques bien connues liées à la perte de positivité et de symétrie de ces matrices. Ces algorithmes robustes sont finalement évalués et comparés dans le cadre d'une application de biométrie de l'iris à partir de vidéos. Le suivi de la pupille est exploitée pour enrôler et identifier en temps-réel une personne grâce à son iris-code. / The family of Kalman filter model allows to estimate the states of a dynamical system from a set of observations. Despite a simple model, these filters are used in a large field of applications: Radar, vision and communications. The success is mainly based on the existence of exact smoothing or filtering algorithms, \ie linear to the number of observations and which minimize the mean square error. In this thesis, we are concerned about the pairwise Kalman filter. This filter adds from the orignal model, new interactions between hidden states and obervations while keeping exact algorithms in the case of linear and Gaussian models. We studied particularly the problem of the unsupervised and robust estimation of a pairwise Kalman filter parameters from a limited set of observations. The manuscript describes several learning algorithms by the estimation of the likelihood maximum according to EM (\textit{Expectation-Maximization}) principle. These original algorithms allow to embed a-priori constraints on studied system parameters, like a knowledge about physical or sensors. These constrained systems reduce the ambiguity, linked to identifiability issue of the pairwise Kalman filter during the parameter estimation. They allow also to limit the number of local maxima of likelihood function with the reduction of the dimension of search space and avoid sometime the trapping of EM algorithm. It is important to note that all proposed algorithm of this manuscrit can be applied to the original Kalman filter, as a particular pairwise Kalman filter. All algorithm are made robust by the propagation of square root matrices instead of the covariance matrices, which allows to limit the numerical issues, linked to the loses of symetry or positivity of these matrices. These algorithm are finally evaluated and compared in the case of an iris biometry application from video sequences. Pupil tracking is used to enroll and recognize in real-time a person thanks to its iris-code.
|
7 |
Mutual Learning Algorithms in Machine LearningChowdhury, Sabrina Tarin 05 1900 (has links)
Indiana University-Purdue University Indianapolis (IUPUI) / Mutual learning algorithm is a machine learning algorithm where multiple machine learning algorithms learns from different sources and then share their knowledge among themselves so that all the agents can improve their classification and prediction accuracies simultaneously. Mutual learning algorithm can be an efficient mechanism for improving the machine learning and neural network efficiency in a multi-agent system. Usually, in knowledge distillation algorithms, a big network plays the role of a static teacher and passes the data to smaller networks, known as student networks, to improve the efficiency of the latter. In this thesis, it is showed that two small networks can dynamically and interchangeably play the changing roles of teacher and student to share their knowledge and hence, the efficiency of both the networks improve simultaneously. This type of dynamic learning mechanism can be very useful in mobile environment where there is resource constraint for training with big dataset. Data exchange in multi agent, teacher-student network system can lead to efficient learning. The concept and the proposed mutual learning algorithm are demonstrated using convolutional neural networks (CNNs) and Support Vector Machine (SVM) to recognize the pattern recognition problem using MNIST hand-writing dataset.
The concept of machine learning is applied in the field of natural language processing (NLP) too. Machines with basic understanding of human language are getting increasingly popular in day-to-day life. Therefore, NLP-enabled machines with memory efficient training can potentially become an indispensable part of our life in near future. A classic problem in the field of NLP is news classification problem where news articles from newspapers are classified by news categories by machine learning algorithms. In this thesis, we show news classification implemented using Naïve Bayes and support vector machine (SVM) algorithm. Then we show two small networks can dynamically play the changing roles of teacher and student to share their knowledge on news classification and hence, the efficiency of both the networks improves simultaneously. The mutual learning algorithm is applied between homogenous algorithms first, i.e., between two Naive Bayes algorithms and two SVM algorithms. Then the mutual learning is demonstrated between heterogenous agents, i.e., between one Naïve Bayes and one SVM agent and the relative efficiency increase between the agents is discussed before and after mutual learning. / 2025-04-04
|
8 |
Application of Reinforcement Learning to Multi-Agent Production SchedulingWang, Yi-chi 13 December 2003 (has links)
Reinforcement learning (RL) has received attention in recent years from agent-based researchers because it can be applied to problems where autonomous agents learn to select proper actions for achieving their goals based on interactions with their environment. Each time an agent performs an action, the environment¡Šs response, as indicated by its new state, is used by the agent to reward or penalize its action. The agent¡Šs goal is to maximize the total amount of reward it receives over the long run. Although there have been several successful examples demonstrating the usefulness of RL, its application to manufacturing systems has not been fully explored. The objective of this research is to develop a set of guidelines for applying the Q-learning algorithm to enable an individual agent to develop a decision making policy for use in agent-based production scheduling applications such as dispatching rule selection and job routing. For the dispatching rule selection problem, a single machine agent employs the Q-learning algorithm to develop a decision-making policy on selecting the appropriate dispatching rule from among three given dispatching rules. In the job routing problem, a simulated job shop system is used for examining the implementation of the Q-learning algorithm for use by job agents when making routing decisions in such an environment. Two factorial experiment designs for studying the settings used to apply Q-learning to the single machine dispatching rule selection problem and the job routing problem are carried out. This study not only investigates the main effects of this Q-learning application but also provides recommendations for factor settings and useful guidelines for future applications of Q-learning to agent-based production scheduling.
|
9 |
Detecting and preventing the electronic transmission of illicit imagesIbrahim, Amin Abdurahman 01 April 2009 (has links)
The sexual exploitation of children remains a very serious problem and is rapidly increasing globally through the use of the Internet. This work focuses on the current methods employed by criminals to generate and distribute child pornography, the methods used by law enforcement agencies to deter them, and the drawbacks of currently used methods, as well as the surrounding legal and privacy issues. A proven method to detect the transmission of illicit images at the network layer is presented within this paper. With this research, it is now possible to actively filter illicit pornographic images as they are transmitted over the network layer in real-time. It is shown that a Stochastic Learning Weak Estimator learning algorithm and a Maximum Likelihood Estimator learning algorithm can be applied against Linear Classifiers to identify and filter illicit pornographic images. In this thesis, these two learning algorithms were combined with algorithms such as the Non-negative Vector Similarity Coefficient-based Distance algorithm, Euclidian Distance, and Weighted Euclidian Distance. Based upon this research, a prototype was developed using the abovementioned system, capable of performing classification on both compressed and uncompressed images. Experimental results showed that classification accuracies and the overhead of network-based approaches did have a significant effect on routing devices. All images used in our experiments were legal. No actual child pornography images were ever collected, seen, sought, or used.
|
10 |
De-quantizing quantum machine learning algorithmsSköldhed, Stefanie January 2022 (has links)
Today, a modern and interesting research area is machine learning. Another new and exciting research area is quantum computation, which is the study of the information processing tasks accomplished by practising quantum mechanical systems. This master thesis will combine both areas, and investigate quantum machine learning. Kerenidis’ and Prakash’s quantum algorithm for recommendation systems, that offered exponential speedup over the best known classical algorithms at the time, will be examined together with Tang’s classical algorithm regarding recommendation systems, which operates in time only polynomial slower than the previously mentioned algorithm. The speedup in the quantum algorithm was achieved by assuming that the algorithm had quantum access to the data structure and that the mapping to the quantum state was performed in polylog(mn). The speedup in the classical algorithm was attained by assuming that the sampling could be performed in O(logn) and O(logmn) for vectors and matrices, respectively.
|
Page generated in 0.091 seconds