Reinforcement Learning by Policy Search

One objective of artificial intelligence is to model the behavior of an intelligent agent interacting with its environment. The environment's transformations can be modeled as a Markov chain, whose state is partially observable to the agent and affected by its actions; such processes are known as partially observable Markov decision processes (POMDPs). While the environment's dynamics are assumed to obey certain rules, the agent does not know them and must learn. In this dissertation we focus on the agent's adaptation as captured by the reinforcement learning framework. This means learning a policy---a mapping of observations into actions---based on feedback from the environment. The learning can be viewed as browsing a set of policies while evaluating them by trial through interaction with the environment. The set of policies is constrained by the architecture of the agent's controller. POMDPs require a controller to have a memory. We investigate controllers with memory, including controllers with external memory, finite state controllers and distributed controllers for multi-agent systems. For these various controllers we work out the details of the algorithms which learn by ascending the gradient of expected cumulative reinforcement. Building on statistical learning theory and experiment design theory, a policy evaluation algorithm is developed for the case of experience re-use. We address the question of sufficient experience for uniform convergence of policy evaluation and obtain sample complexity bounds for various estimators. Finally, we demonstrate the performance of the proposed algorithms on several domains, the most complex of which is simulated adaptive packet routing in a telecommunication network.

Safe Distributed Coordination of Heterogeneous Robots through Dynamic Simple Temporal Networks

Research on autonomous intelligent systems has focused on how robots can robustly carry out missions in uncertain and harsh environments with very little or no human intervention. Robotic execution languages such as RAPs, ESL, and TDL improve robustness by managing functionally redundant procedures for achieving goals. The model-based programming approach extends this by guaranteeing correctness of execution through pre-planning of non-deterministic timed threads of activities. Executing model-based programs effectively on distributed autonomous platforms requires distributing this pre-planning process. This thesis presents a distributed planner for modelbased programs whose planning and execution is distributed among agents with widely varying levels of processor power and memory resources. We make two key contributions. First, we reformulate a model-based program, which describes cooperative activities, into a hierarchical dynamic simple temporal network. This enables efficient distributed coordination of robots and supports deployment on heterogeneous robots. Second, we introduce a distributed temporal planner, called DTP, which solves hierarchical dynamic simple temporal networks with the assistance of the distributed Bellman-Ford shortest path algorithm. The implementation of DTP has been demonstrated successfully on a wide range of randomly generated examples and on a pursuer-evader challenge problem in simulation.

Factorial Hidden Markov Models

We present a framework for learning in hidden Markov models with distributed state representations. Within this framework, we derive a learning algorithm based on the Expectation--Maximization (EM) procedure for maximum likelihood estimation. Analogous to the standard Baum-Welch update rules, the M-step of our algorithm is exact and can be solved analytically. However, due to the combinatorial nature of the hidden state representation, the exact E-step is intractable. A simple and tractable mean field approximation is derived. Empirical results on a set of problems suggest that both the mean field approximation and Gibbs sampling are viable alternatives to the computationally expensive exact algorithm.

On the Convergence of Stochastic Iterative Dynamic Programming Algorithms

Recent developments in the area of reinforcement learning have yielded a number of new algorithms for the prediction and control of Markovian environments. These algorithms, including the TD(lambda) algorithm of Sutton (1988) and the Q-learning algorithm of Watkins (1989), can be motivated heuristically as approximations to dynamic programming (DP). In this paper we provide a rigorous proof of convergence of these DP-based learning algorithms by relating them to the powerful techniques of stochastic approximation theory via a new convergence theorem. The theorem establishes a general class of convergent algorithms to which both TD(lambda) and Q-learning belong.

Enhancing fault management performance of two-step QoS routing algorithms in GMPLS networks

In this paper a novel methodology aimed at minimizing the probability of network failure and the failure impact (in terms of QoS degradation) while optimizing the resource consumption is introduced. A detailed study of MPLS recovery techniques and their GMPLS extensions are also presented. In this scenario, some features for reducing the failure impact and offering minimum failure probabilities at the same time are also analyzed. Novel two-step routing algorithms using this methodology are proposed. Results show that these methods offer high protection levels with optimal resource consumption

Enhancing MPLS QoS routing algorithms by using the network protection degree paradigm

IP based networks still do not have the required degree of reliability required by new multimedia services, achieving such reliability will be crucial in the success or failure of the new Internet generation. Most of existing schemes for QoS routing do not take into consideration parameters concerning the quality of the protection, such as packet loss or restoration time. In this paper, we define a new paradigm to develop new protection strategies for building reliable MPLS networks, based on what we have called the network protection degree (NPD). This NPD consists of an a priori evaluation, the failure sensibility degree (FSD), which provides the failure probability and an a posteriori evaluation, the failure impact degree (FID), to determine the impact on the network in case of failure. Having mathematical formulated these components, we point out the most relevant components. Experimental results demonstrate the benefits of the utilization of the NPD, when used to enhance some current QoS routing algorithms to offer a certain degree of protection

A system to evaluate the accuracy of a visual mosaicking methodology

When underwater vehicles navigate close to the ocean floor, computer vision techniques can be applied to obtain motion estimates. A complete system to create visual mosaics of the seabed is described in this paper. Unfortunately, the accuracy of the constructed mosaic is difficult to evaluate. The use of a laboratory setup to obtain an accurate error measurement is proposed. The system consists on a robot arm carrying a downward looking camera. A pattern formed by a white background and a matrix of black dots uniformly distributed along the surveyed scene is used to find the exact image registration parameters. When the robot executes a trajectory (simulating the motion of a submersible), an image sequence is acquired by the camera. The estimated motion computed from the encoders of the robot is refined by detecting, to subpixel accuracy, the black dots of the image sequence, and computing the 2D projective transform which relates two consecutive images. The pattern is then substituted by a poster of the sea floor and the trajectory is executed again, acquiring the image sequence used to test the accuracy of the mosaicking system

Multicast algorithms evaluation using an adaptative routing in ATM networks

This paper proposes a multicast implementation based on adaptive routing with anticipated calculation. Three different cost measures for a point-to-multipoint connection: bandwidth cost, connection establishment cost and switching cost can be considered. The application of the method based on pre-evaluated routing tables makes possible the reduction of bandwidth cost and connection establishment cost individually

Automatic self-configuration of the logical network using distributed software agents

We present a system for dynamic network resource configuration in environments with bandwidth reservation. The proposed system is completely distributed and automates the mechanisms for adapting the logical network to the offered load. The system is able to manage dynamically a logical network such as a virtual path network in ATM or a label switched path network in MPLS or GMPLS. The system design and implementation is based on a multi-agent system (MAS) which make the decisions of when and how to change a logical path. Despite the lack of a centralised global network view, results show that MAS manages the network resources effectively, reducing the connection blocking probability and, therefore, achieving better utilisation of network resources. We also include details of its architecture and implementation

Improving Clustering Algorithms for Image Segmentation using Contour and Region Information

In image segmentation, clustering algorithms are very popular because they are intuitive and, some of them, easy to implement. For instance, the k-means is one of the most used in the literature, and many authors successfully compare their new proposal with the results achieved by the k-means. However, it is well known that clustering image segmentation has many problems. For instance, the number of regions of the image has to be known a priori, as well as different initial seed placement (initial clusters) could produce different segmentation results. Most of these algorithms could be slightly improved by considering the coordinates of the image as features in the clustering process (to take spatial region information into account). In this paper we propose a significant improvement of clustering algorithms for image segmentation. The method is qualitatively and quantitative evaluated over a set of synthetic and real images, and compared with classical clustering approaches. Results demonstrate the validity of this new approach

A distributed environment for virtual and/or real experiments for underwater robots

This paper presents the distributed environment for virtual and/or real experiments for underwater robots (DEVRE). This environment is composed of a set of processes running on a local area network composed of three sites: 1) the onboard AUV computer; 2) a surface computer used as human-machine interface (HMI); and 3) a computer used for simulating the vehicle dynamics and representing the virtual world. The HMI can be transparently linked to the real sensors and actuators dealing with a real mission. It can also be linked with virtual sensors and virtual actuators, dealing with a virtual mission. The aim of DEVRE is to assist engineers during the software development and testing in the lab prior to real experiments

Dynamic bandwidth management as part of an integrated network management system based on distributed agents

We present a system for dynamic network resource configuration in environments with bandwidth reservation and path restoration mechanisms. Our focus is on the dynamic bandwidth management results, although the main goal of the system is the integration of the different mechanisms that manage the reserved paths (bandwidth, restoration, and spare capacity planning). The objective is to avoid conflicts between these mechanisms. The system is able to dynamically manage a logical network such as a virtual path network in ATM or a label switch path network in MPLS. This system has been designed to be modular in the sense that in can be activated or deactivated, and it can be applied only in a sub-network. The system design and implementation is based on a multi-agent system (MAS). We also included details of its architecture and implementation

Sistema distribuido colaborativo para la recuperación de documentación educativa.

Sobre la base del estudio de esquemas de comunicación asincrónicos, el presente artículo propone un framework basado en servicios web para la recuperación de documentación educativa en forma colaborativa en ambientes distribuidos, y más concretamente 1) propone describir explícitamente la intención de los mensajes entre procesos utilizando para ello performatives del estándar FIPA, y sobre la base de dicha descripción 2) presenta un conjunto de métricas que evalúan la gestión de la información y 3) define dos algoritmos que mejoran la calidad del servicio del intercambio de información en función de la mejora del proceso de recuperación de documentación educativa.

A Focus on Distributed Denial of Service

An Arbor Networks paper describing DDoS attacks and related attacks. The first 9-10 pages or so are good background reading for INFO6003. Students may also find the rest of the paper interesting.