Adaptive learning of symbolic numerical constraints in the real-world

This thesis presents an adaptive learning system which can learn symbolic hypotheses representing numerical constraints using the observations of a robot. The system is based on a framework which deals with real-world conditions of noise and concept drift. Inductive Logic Programming (ILP) is used as the base learning method. ILP is a machine learning approach that uses first-order logic (FOL) for knowledge representation. FOL representation allows expressing relational features, which is not possible with attribute-based machine learning methods. Another advantage of using FOL is that it allows reasoning to derive facts from the given observations and background knowledge. FOL is also used in planning and reasoning systems, which makes the learned rules easily adaptable to these systems. The core learning method extends a well-known ILP system with a constraint solver and lazy evaluation. This extension deals with the limitation of ILP in learning numerical constraints. The extended learning method imposes constraints on the domains of the numerical variables in a hypothesis clause. If an appropriate constraint exists, addition of it transforms a previously inconsistent hypothesis into a useful one. Hence, the extended method combines the relational learning capability of Inductive Logic Programming and the numerical reasoning capability of Constraint Logic Programming. Finding appropriate constraints on the domains of variables is achieved using a constraint solver. During the ILP hypothesis generation, if a hypothesis contains a numerical variable, then a constraint satisfaction problem (CSP) is solved to find the largest interval in the domain of the variable, such that, when the values in this interval are substituted with this variable in the hypothesis, the hypothesis does not entail any of the negative examples but it entails at least one positive example. If such an interval exists, constraining the domain of this variable makes the hypothesis consistent. Since robot applications are targeted, learning symbolic numerical constraints should be robust under real-world conditions. The proposed system integrates a noise removal module to cope with uncertainties arising from the robot and its environment. Local outlier factor (LOF) method is used for noise removal since it does not require a prior cluster scheme and it takes densities into account. Noise removal is applied on the inputs of the constraint solver since it does not handle the noise itself. Another addition is a concept drift detector which makes the system adaptive to external and internal changes. A concept drift may render the previously learned hypotheses obsolete. Hence, the drift detection module continuously monitors the prediction success of the learned model, and requests the update of the learned hypotheses if necessary. The system is evaluated with robot experiments in the real-world and computer generated scenarios. The results of the experiments show that the enhancements increase the robustness and the effectiveness of the learner. It is observed that the presented system can learn from noisy data that is collected in the real-world environment, and it can improve the prediction performance by detecting concept drifts. The system is expected to be useful in lifelong learning scenarios and in compromising between low-level and high-level robot learning.