Cristina Tîrnăucă

Behavioral Modeling Based on Probabilistic Finite Automata: An Empirical Study

Imagine an agent that performs tasks according to different strategies. The goal of Behavioral Recognition (BR) is to identify which of the available strategies is the one being used by the agent, by simply observing the agent’s actions and the environmental conditions during a certain period of time. The goal of Behavioral Cloning (BC) is more ambitious. In this last case, the learner must be able to build a model of the behavior of the agent. In both settings, the only assumption is that the learner has access to a training set that contains instances of observed behavioral traces for each available strategy. This paper studies a machine learning approach based on Probabilistic Finite Automata (PFAs), capable of achieving both the recognition and cloning tasks. We evaluate the performance of PFAs in the context of a simulated learning environment (in this case, a virtual Roomba vacuum cleaner robot), and compare it with a collection of other machine learning approaches.

Border algorithms for computing hasse diagrams of arbitrary lattices

The Border algorithm and the iPred algorithm find the Hasse diagrams of FCA lattices. We show that they can be generalized to arbitrary lattices. In the case of iPred, this requires the identification of a join-semilattice homomorphism into a distributive lattice.

Automatic Generation of User Interaction Models

A prominent requirement in the field of human-computer interaction is to make mobile applications more usable and better adjusted to their users’ needs. In particular, designers of groupware applications face the task of developing software for many users while making it work as if it was designed for each single individual. User modeling research has attempted to address these issues. A precondition for achieving this task is to find predictive and generative models of the user interactions. In this paper we develop a methodology for modeling the user behavior when interacting with a computer system. The byproduct of this methodology is a low level representation of the user interactions in the form of weighted automata, which can be easily transformed into user profiles in text form. Profiles can then be used by the designer to configure and verify the task model of the system.

Model-driven regularization approach to straight line program genetic programming

A regularization method for linear genetic programming is proposed.Straight line programs with transcendental elementary functions are used.A sharp bound for the Vapnik-Chervonenkis dimension of programs is encountered.Our approach is empirically better than other statistical regularization methods. This paper presents a regularization method for program complexity control of linear genetic programming tuned for transcendental elementary functions. Our goal is to improve the performance of evolutionary methods when solving symbolic regression tasks involving Pfaffian functions such as polynomials, analytic algebraic and transcendental operations like sigmoid, inverse trigonometric and radial basis functions. We propose the use of straight line programs as the underlying structure for representing symbolic expressions. Our main result is a sharp upper bound for the Vapnik Chervonenkis dimension of families of straight line programs containing transcendental elementary functions. This bound leads to a penalization criterion for the mean square error based fitness function often used in genetic programming for solving inductive learning problems. Our experiments show that the new fitness function gives very good results when compared with classical statistical regularization methods (such as Akaike and Bayesian Information Criteria) in almost all studied situations, including some benchmark real-world regression problems.

Global optimality in k-means clustering

Abstract We study the problem of finding an optimum clustering, a problem known to be NP-hard. Existing literature contains algorithms running in time proportional to the number of points raised to a power that depends on the dimensionality and on the number of clusters. Published validations of some of these algorithms are unfortunately incomplete; besides, the constant factors (with respect to the number of points) in their running time bounds have seen several published important improvements but are still huge, exponential on the dimension and on the number of clusters, making the corresponding algorithms fully impractical. We provide a new algorithm, with its corresponding complexity-theoretic analysis. It reduces both the exponent and the constant factor, to the extent that it becomes feasible for relevant particular cases. Additionally, it parallelizes extremely well, so that its implementation on current high-performance hardware is quite straightforward. Our proposal opens the door to potential improvements along a research line that had no practical significance so far; besides, a long but single-shot run of our algorithm allows one to identify absolutely optimum solutions for benchmark problems, whereby alternative heuristic proposals can evaluate the goodness of their solutions and the precise price paid for their faster running times.

Daily Routines Inference Based on Location History.

The huge amount of location tracker data generated by electronic devices makes them an ideal source of information for detecting trends and behaviors in their users’ lives. Learning these patterns is very important for recommender systems or applications targeted at behavior prediction. In this work we show how user location history can be processed in order to extract the most relevant visited locations and to model the user’s profile through a weighted finite automaton, a probabilistic graphical structure that is able to handle locations and temporal context compactly. Our condensed representation gives access to the user’s routines and can play an important role in recommender systems.

User Interaction Modeling and Profile Extraction in Interactive Systems: A Groupware Application Case Study

A relevant goal in human–computer interaction is to produce applications that are easy to use and well-adjusted to their users’ needs. To address this problem it is important to know how users interact with the system. This work constitutes a methodological contribution capable of identifying the context of use in which users perform interactions with a groupware application (synchronous or asynchronous) and provides, using machine learning techniques, generative models of how users behave. Additionally, these models are transformed into a text that describes in natural language the main characteristics of the interaction of the users with the system.

Teaching a Virtual Robot to Perform Tasks by Learning from Observation

We propose a methodology based on Learning from Observation in order to teach a virtual robot to perform its tasks. Our technique only assumes that behaviors to be cloned can be observed and represented using a finite alphabet of symbols. A virtual agent is used to generate training material, according to a range of strategies of gradually increasing complexity. We use Machine Learning techniques to learn new strategies by observing and thereafter imitating the actions performed by the agent. We perform several experiments to test our proposal. The analysis of those experiments suggests that probabilistic finite state machines could be a suitable tool for the problem of behavioral cloning. We believe that the given methodology is easy to integrate in the learning module of any Ubiquitous Robot Architecture.

Types of trusted information that make DFA identification with correction queries feasible

In the query learning model, the problem of efficiently identifying a deterministic finite automaton (DFA) has been widely investigated. While DFAs are known to be polynomial time learnable with a combination of membership queries (MQs) and equivalence queries (EQs), each of these types of queries alone are not enough to provide sufficient information for the learner. Therefore, the possibility of having some extra-information shared between the learner and the teacher has been discussed. In this paper, the problem of efficient DFA identification with correction queries (CQs) - an extension of MQs - when additional information is provided to the learner is addressed. We show that knowing the number of states of the target DFA does not help (similar to the case of MQs or EQs), but other parameters such as the reversibility or injectivity degree are useful.