Abstraction in Control Learning
Richard C. Yee · 1992
Efficient learning requires strong generalization biases, which are often provided to learning agents through carefully designed input representations. Human novices, however, are capable of using experience in a task to develop their own expert representations. This report is a dissertation proposal examining how agents can acquire learning biases through the construction of abstract representations. Efficient learning and planning call for representing specific experiences in terms of abstract features and concepts that reflect the goals and dynamics of tasks. Recent AI research has begun to incorporate established results from control theory, which studies the interaction between agents and dynamical systems or environments. Learning methods related to dynamic programming (DP) address the problem of ...