Learning structured task related abstractions
Penkov, Svetlin Valentinov
As robots and autonomous agents are to assist people with more tasks in various domains they need the ability to quickly gain contextual awareness in unseen environments and learn new tasks. Current state of the art methods rely predominantly on statistical learning techniques which tend to overfit to sensory signals and often fail to extract structured task related abstractions. The obtained environment and task models are typically represented as black box objects that cannot be easily updated or inspected and provide limited generalisation capabilities. We address the aforementioned shortcomings of current methods by explicitly studying the problem of learning structured task related abstractions. In particular, we are interested in extracting symbolic representations of the environment from sensory signals and encoding the task to be executed as a computer program. We consider the standard problem of learning to solve a task by mapping sensory signals to actions and propose the decomposition of such a mapping into two stages: i) perceiving symbols from sensory data and ii) using a program to manipulate those symbols in order to make decisions. This thesis studies the bidirectional interactions between the agent’s capabilities to perceive symbols and the programs it can execute in order to solve a task. In the first part of the thesis we demonstrate that access to a programmatic description of the task provides a strong inductive bias which facilitates the learning of structured task related representations of the environment. In order to do so, we first consider a collaborative human-robot interaction setup and propose a framework for Grounding and Learning Instances through Demonstration and Eye tracking (GLIDE) which enables robots to learn symbolic representations of the environment from few demonstrations. In order to relax the constraints on the task encoding program which GLIDE assumes, we introduce the perceptor gradients algorithm and prove that it can be applied with any task encoding program. In the second part of the thesis we investigate the complement problem of inducing task encoding programs assuming that a symbolic representations of the environment is available. Therefore, we propose the p-machine – a novel program induction framework which combines standard enumerative search techniques with a stochastic gradient descent optimiser in order to obtain an efficient program synthesiser. We show that the induction of task encoding programs is applicable to various problems such as learning physics laws, inspecting neural networks and learning in human-robot interaction setups.