Tesis de maestría
Probabilistic Agent Localization and Fuzzy-Bayesian Pass Evaluation for the RoboCup Simulation 3D League-Edición Única
Fecha
2007-12-01Autor
Bustamante Horta, Carlos Fernando
Institución
Resumen
In most real-time multiagent systems, the environment tends to be very complex,
partially observable, stochastic, sequential, dynamic and continuous. In such constantly
changing and challenging environments, agents must be prepared to deal with
uncertainty and noisy measurements when making decisions.
This thesis aims at describing, analyzing, implementing and discussing probabilistic
and statistical machine learning for action evaluation in the RoboCup 3D Soccer
Simulation domain. Many of the issues appear when developing a RoboCup 3D team.
Soccer is a game that presents a really complex environment like the one described
above. This makes it a really good testbed for machine learning algorithms.
In order to successfully apply machine learning for evaluating actions, a RoboCup
3D agent needs to have a good and accurate world model. That is why this thesis
divides the problem in three layers: obtaining the motion models, self localization and
classification for action evaluation. An entire chapter is dedicated to the first layer for
obtaining the motion models of the spherical agents and the ball. This allows the agent
to understand and predict its actions and precisely localize itself in the environment.
Another chapter is dedicated for the second layer, explaining and testing an interesting
and innovative approach to mobile robot localization using particle filters called
Monte Carlo Localization (MCL). The motion models from the first layer are used in
MCL to sample particles in the prediction phase of the particle filter. Furthermore,
the filter is enhanced using injection of random particles with a proposed Kalman filter
sensor fusion strategy.
The third layer is covered in another chapter, using classification for action evaluation.
Classification, or concept learning, is a special type of supervised machine
learning in which a function is estimated from a training set of labeled attribute vectors
or examples. Thus, when a new example arrives, it is labeled using the learned
function, i.e. classified as a specific concept. A commonly known classifier that deals
with uncertainty in computationally tractable time and space is the Naive Bayes classifier.
Although its conditional independence assumption is not always accomplished,
it has proved to be successful in a whole range of applications. Typically, the Naive
Bayes model assumes discrete attributes, but modifications have been proposed in the
literature for handling in continuous domains. This chapter analyzes the performance
of a hybrid Fuzzy Naive Bayes classifier in the RoboCup Simulation 3D domain in
which attributes are treated as fuzzy variables. A comparison is made between the
aforementioned approach and a Gaussian Naive Bayes classifier using the pass skill as
a testbed.
Results show that the motion models were precise when evaluating the parameters
with noiseless data. The hybrid localization method works well even for global localization
problems in the sense that it recovers relatively fast from errors and is accurate.
On the other hand, the hybrid action evaluation mechanism gave acceptable results
with approximately 80% of the examples correclty classified in a test set for a pass skill
scenario. The last chapter shows that there is a lot of future research on these topics.