Archive for May, 1996

Introspective Multistrategy Learning: Constructing a Learning Strategy under Reasoning Failure

The thesis put forth by this dissertation is that introspective analyses facilitate the construction of learning strategies. Furthermore, learning is much like nonlinear planning and problem solving. Like problem solving, it can be specified by a set of explicit learning goals (i.e., desired changes to the reasoner’s knowledge); these goals can be achieved by constructing a plan from a set of operators (the learning algorithms) that execute in a knowledge space. However, in order to specify learning goals and to avoid negative interactions between operators, a reasoner requires a model of its reasoning processes and knowledge.

With such a model, the reasoner can declaratively represent the events and causal relations of its mental world in the same manner that it represents events and relations in the physical world. This representation enables introspective self-examination, which contributes to learning by providing a basis for identifying what needs to be learned when reasoning fails. A multistrategy system possessing several learning algorithms can decide what to learn, and which algorithm(s) to apply, by analyzing the model of its reasoning. This introspective analysis therefore allows the learner to understand its reasoning failures, to determine the causes of the failures, to identify needed knowledge repairs to avoid such failures in the future, and to build a learning strategy (plan).

Thus, the research goal is to develop both a content theory and a process theory of introspective multistrategy learning and to establish the conditions under which such an approach is fruitful. Empirical experiments provide results that support the claims herein. The theory was implemented in a computational model called Meta-AQUA that attempts to understand simple stories. The system uses case-based reasoning to explain reasoning failures and to generate sets of learning goals, and it uses a standard non-linear planner to achieve these goals.

Evaluating Meta-AQUA with and without learning goals generated results indicating that computational introspection facilitates the learning process. In particular, the results lead to the conclusion that the stage that posts learning goals is a necessary stage if negative interactions between learning methods are to be avoided and if learning is to remain effective.

Read the thesis:

Introspective multistrategy learning: Constructing a learning strategy under reasoning failure

by Michael T. Cox

PhD Thesis, Technical Report GIT-CC-96/06, College of Computing, Georgia Institute of Technology, Atlanta, GA, 1996

Learning Adaptive Reactive Agents

An autonomous agent is an intelligent system that has an ongoing interaction with a dynamic external world. It can perceive and act on the world through a set of limited sensors and effectors. Its most important characteristic is that it is forced to make decisions sequentially, one after another, during its entire “life”. The main objective of this dissertation is to study algorithms by which an autonomous agents can learn, using their own experience, to perform sequential decision-making efficiently and autonomously. The dissertation describes a framework for studying autonomous sequential decision-making consisting of three main elements: the agent, the environment, and the task. The agent attempts to control the environment by perceiving the environment and choosing actions in a sequential fashion. The environment is a dynamic system characterized by a state and its dynamics, a function that describes the evolution of the state given the agent’s actions. A task is a declarative description of the desired behavior the agent should exhibit as it interacts with the environment. The ultimate goal of the agent is to learn a policy or strategy for selecting actions that maximizes its expected benefit as defined by the task.

The dissertation focuses on sequential decision-making when the environment is characterized by continuous states and actions, and the agent has imperfect perception, incomplete knowledge, and limited computational resources. The main characteristic of the approach proposed in this dissertation is that the agent uses its previous experiences to improve estimates of the long-term benefit associated with the execution of specific actions. The agent uses these estimates to evaluate how desirable is to execute alternative actions and select the one that best balances the short- and long-term consequences, taking special consideration of the expected benefit associated with actions that accomplish new learning while making progress on the task.

The approach is based on novel methods that are specifically designed to address the problems associated with continuous domains, imperfect perception, incomplete knowledge, and limited computational resources. The approach is implemented using case-based techniques and extensively evaluated in simulated and real systems including autonomous mobile robots, pendulum swinging and balancing controllers, and other non-linear dynamic system controllers.

Read the thesis:

Learning Adaptive Reactive Agents

by Juan Carlos Santamaria

PhD Thesis, College of Computing, Georgia Institute of Technology, Atlanta, GA, 1996