| The general subject matter of this thesis pertains to the study of the framework within which one can construct, and experiment with intelligent adaptive systems. One of the important contributions of this thesis is the generalization of the feedback loops, currently studied in the fields of Reinforcement Learning and Learning Automata, to allow for a wide variety of experiments, including multi-agent and multi-environment interactions. This generalization naturally led to the investigation of learning to play games under extreme conditions of minimal a priori information. As another outcome of the above generalization, we consider alternatives to the reward/punishment signals for supplying the goal information to the artificial systems. This work has resulted in other important contributions---the classification of goals according to their arity, and the subsequent study of the novel feedback signal as an alternative to reinforcement. A learning task can now be analyzed and classified based on its arity. The introduction of the feedback signal has further led to the development of the discretized forms of learning algorithms and to DQ-learning in particular. Knowledge of the arity of a task allows us to construct the corresponding feedback signal, and, hence, to apply the DQ-learning algorithm with guaranteed convergence and memory savings which enable us to tackle more challenging tasks. Given that these ideas were borne out of our generalization of the agent-environment interaction; we believe that we have shown this generalization to be fruitful. In particular, we have demonstrated its theoretical merit and its applicability in an ensemble of areas including game playing, organization of data in adaptive lists, and differentiated robot control. We further believe that other important contributions can result from this work. |
