This thesis demonstrates how the power of symbolic processing can be exploited in the learning of low level control functions. It proposes a novel hybrid architecture with a tight coupling between a variant of symbolic planning and reinforcement learning. This architecture combines the strengths of the function approximation of subsymbolic learning with the more abstract compositional nature of symbolic learning. The former is able to represent mappings of world states to actions in an accurate way. The latter allows a more rapid solution to problems by exploiting structure within the domain. A control function is learnt over time through interaction with the world. Symbols are attached to features in the functions. The symbolic attachments act as anchor points used to transform the function of a previously learnt task to that of a new task. The solution of more complex tasks is achieved through composing simpler functions, using the symbolic attachments to determine the composition. T...

Often the best model to solve a real-world problem is relatively complex. This paper presents oracle learning, a method using a larger model as an oracle to train a smaller model on unlabeled data in order to obtain (1) a smaller acceptable model and (2) improved results over standard training methods on a similarly sized smaller model. In particular, this paper looks at oracle learning as applied to multi-layer perceptrons trained using standard backpropagation. Using multi-layer perceptrons for both the larger and smaller models, oracle learning obtains a 15.16 % average decrease in error over direct training while retaining 99.64 % of the initial oracle accuracy on automatic spoken digit recognition with networks on average only 7 % of the original size. For optical character recognition, oracle learning results in neural networks 6 % of the original size that yield a 11.40 % average decrease in error over direct training while maintaining 98.95 % of the initial oracle accuracy. 1

Abstract. Effective backpropagation training of multi-layer perceptrons depends on the incorporation of an appropriate error or objective function. Classification-based (CB) error functions are heuristic approaches that attempt to guide the network directly to correct pattern classification rather than using common error minimization heuristics, such as sum-squared error and cross-entropy, which do not explicitly minimize classification error. This work presents CB3, a novel CB approach that learns the error function to be used while training. This is accomplished by learning pattern confidence margins during training, which are used to dynamically set output target values for each training pattern. On eleven applications, CB3 significantly outperforms previous CB error functions, and also reduces average test error over conventional error metrics using 0-1 targets without weight decay by 1.8%, and by 1.3 % over metrics with weight decay. CB3 also exhibits lower model variance and tighter mean confidence interval.

Empirical studies of multitask learning provide some evidence that the performance of a learning system on its intended targets improves by presenting to the learning system additional related tasks, also called contexts, as additional input. Angluin, Gasarch, and Smith, as well as Kinber, Smith, Velauthapillai, and Wiehagen have provided mathematical justification for this phenomenon in the inductive inference framework. However, their proofs rely heavily on self-referential coding tricks, that is, they directly code the solution of the learning problem into the context. In this work we prove, in the inductive inference setting, that multitask learning is extremely powerful even without using obvious coding tricks. Coding tricks are avoided in the powerful sense of Fulk's notion of robust learning. Also, studied is the difficulty of the functional dependence between the intended target tasks and useful associated contexts. Department of CIS, University of Delaware, Newark, DE 19716,...

It is desirable for mobile robots to learn incrementally and to adapt to changes in the environment during their entire lifetime. We propose a method for multi-task learning in reinforcement environments which is applied to a mobile robot learning a navigation task in different environments. We argue that

This work is concerned with practical issues surrounding the application of reinforcement learning to a mobile robot. The robot's task is to navigate in a controlled environment and to collect objects using its gripper. Our aim is to build a control system that enables the robot to learn incrementally and to adapt to changes in the environment. The former is known as multi-task learning, the latter is usually referred to as continual `lifelong' learning. First, we emphasize the connection between adaptive state-space quantisation and continual learning. Second, we describe a novel method for multi-task learning in reinforcement environments. This method is based

of a thesis submitted by