Paola Baldassarri

Navigation with memory in a partially observable environment

Abstract The paper presents an architecture that allows the reactive visual navigation via an unsupervised reinforcement learning. This objective is reached using Q -learning and a hierarchical approach to the developed architecture. Using these techniques requires a deviation from the Partially Observable Markov Decision Processes (POMDP) and some innovations: heuristic techniques for generalizing the experience and for treating the partial observability; a technique for the speed adjournment of the Q function; the definition of a special reinforcement policy adequate for learning a complex task without supervision. The result is a satisfactory learning of the navigation assignment in a simulated environment.

Self-Organizing Maps versus Growing Neural Gas in a Robotic Application

The paper proposes a method for visual based self-localisation of a mobile agent in indoor environment. The images acquired by the camera constitute an implicit topological representation of the environment. The environment is a priori unknown and so the implemented architecture is entirely unsupervised. To compare the performance of some self-organising neural networks, a similar neural network architecture of both Self-Organizing Map (SOM) and Growing Neural Gas (GNG) has been realized. Extensive simulations are provided to characterise the effectiveness of the GNG model in recognition speed, classification tasks and in particular topology preserving as compared to the SOM model. This behaviour depends on the following fact: a network (GNG) that adds nodes into map space can approximate the input space more accurately than a network with a predefined structure and size (SOM). The work shows that the GNG network is able to correctly reconstruct the environment topological map.

Fingerprints Recognition Using Minutiae Extraction: a Fuzzy Approach.

The aim of this paper is to study the fingerprint verification based on local ridge discontinuities features (minutiae) only using grey scale images. We extract minutiae using two algorithms those following ridge lines and then recording ridge endings and bifurcations. Moreover we use a third algorithm able to develop a minutiae verification processing a local area using a neural network ( multilayer perceptron). Fingerprint distortion is filtered using a minutiae whole representation based on regular invariant moments. The results of the three minutiae extraction algorithms are joined during the minutiae pattern matching phase for fingerprint verification. Here we propose a new method of matching that use fuzzy operator to bypass the problem of different numbers of minutiae extracted from the algorithms. Experimental evidences show fingerprint recognition up to 95%.

Multiple Neural Networks System for Dynamic Environments

We propose a “Multiple Neural Networks” system for dynamic environments, where one or more neural nets may no longer be able to properly operate, due to sensible partial changes in the characteristics of the individuals. We assume that each expert network has a reliability factor that can be dynamically re-evaluated on the ground of the global recognition operated by the overall group. Since the net’s “degree of reliability” is defined as “the probability that the net is giving the desired output”, in case of conflicts between the outputs of the various nets the re-evaluation of their “degrees of reliability” can be simply performed on the basis of the Bayes Rule. The new vector of reliability will be used for making the final choice, by applying the “Inclusion based” algorithm over all the maximally consistent subsets of the global outcome. Finally, the nets recognized as responsible for the conflicts will be automatically forced to learn about the changes in the individuals’ characteristics and avoid to make the same error in the immediate future.

Detecting Anomalous Traffic Using Statistical Discriminator and Neural Decisional Motor

One of the main challenges in the information security concerns the introduction of systems able to identify intrusions. In this ambit this work takes place describing a new Intrusion Detection System based on anomaly approach. We realized a system with a hybrid solution between host-based and network-based approaches, and it consisted of two subsystems: a statistical system and a neural one. The features extracted from the network traffic belong only to the IP Header and their trend allows us detecting through a simple visual inspection if an attack occurred. Really the two-tier neural system has to indicate the status of the system. It classifies the traffic of the monitored host, distinguishing the background traffic from the anomalous one. Besides, a very important aspect is that the system is able to classify different instances of the same attack in the same class, establishing which attack occurs.

Hybrid system for a never-ending unsupervised learning

We propose a Hybrid System for dynamic environments, where a “Multiple Neural Networks” system works with Bayes Rule. One or more neural nets may no longer be able to properly operate, due to partial changes in some of the characteristics of the individuals. We assume that each expert network has a reliability factor that can be dynamically re-evaluated on the ground of the global recognition operated by the overall group. Since the nets degree of reliability is defined as the probability that the net is giving the desired output, in case of conflicts between the outputs of the various nets the re-evaluation of their degrees of reliability can be simply performed on the basis of the Bayes Rule. The new vector of reliability will be used for making the final choice, by applying two algorithms, the “Inclusion based” and the “Weighted” one over all the maximally consistent subsets of the global outcome.

Multiple Neural Networks and Bayesian Belief Revision for a never-ending unsupervised learning

A system of Multiple Neural Networks has been proposed to solve the face recognition problem. Our idea is that a set of expert networks specialized to recognize specific parts of face are better than a single network. This is because a single network could no longer be able to correctly recognize the subject when some characteristics partially change. For this purpose we assume that each network has a reliability factor defined as the probability that the network is giving the desired output. In case of conflicts between the outputs of the networks the reliability factor can be dynamically re-evaluated on the base of the Bayes Rrule. The new reliabilities will be used to establish who is the subject. Moreover the network disagreed with the group and specialized to recognize the changed characteristic of the subject will be retrained and then forced to correctly recognize the subject. Then the system is subjected to continuous learning.

An hybrid system for continuous learning

We propose a Multiple Neural Networks system for dynamic environments, where one or more neural nets could no longer be able to properly operate, due to partial changes in some of the characteristics of the individuals We assume that each expert network has a reliability factor that can be dynamically re-evaluated on the ground of the global recognition operated by the overall group Since the nets degree of reliability is defined as the probability that the net is giving the desired output, in case of conflicts between the outputs of the various nets the re-evaluation of their degrees of reliability can be simply performed on the basis of the Bayes Rule The new vector of reliability will be used for making the final choice, by applying two algorithms, the Inclusion based and the Weighted one over all the maximally consistent subsets of the global outcome.

Analysis of Fingerprints Through a Reactive Agent

The aim of this job is to study the process of self-organisation of the knowledge in a reactive autonomous agent that navigates throughout a fingerprint image. This fingerprint has been recorded using a low cost sensor, so it has with her a lot of noise. In this particular situation the usual methods of analysis of the minutiae fail or need a strong pre-processing of the image. Our system is a reactive agent that acts independently from the noise in the image because the process of self-organising of the knowledge carries to the emergency of the concept of “run toward the minutiae” through a categorisation of the sensorial input and a generalisation of the situation “state-action”. The system is based on hybrid architecture for the configuration recognition and the knowledge codifies.

Visual self-localisation using automatic topology construction

The paper proposes a machine learning method for self-localising a mobile agent, using the images supplied by a single omni-directional camera. The images acquired by the camera may be viewed as an implicit topological representation of the environment. The environment is a priori unknown and the topological representation is derived by unsupervised neural network architecture. The architecture includes a self-organising neural network, and is constituted by a growing neural gas, which is well known for its topology preserving quality. The growth depends on the topology that is not a priori defined, and on the need of discovering it, by the neural network, during the learning. The implemented system is able to recognise correctly the input frames and to reconstruct a topological map of the environment. Each node of the neural network identifies a single zone of the environment and the connections between the nodes correspond to the real space connections in the environment.