This paper proposes a robot behavioral μ-selection method that maximizes a quantitative measure of languages in the discrete-event setting. This approach complements Q-learning (also called reinforcement learning) that has been widely used in behavioral robotics to learn primitive behaviors. While μ-selection assigns positive and negative weights to the marked states of a deterministic finite-state automaton (DFSA) model of robot operations, Q-learning as-signs reward/penalty on each transition. While the complexity of Q-learning increases exponentially in the number of states and actions, complexity of μ-selection is polynomial in the number of DFSA states. The paper also presents results of simulation experiments for a robotic scenario to demonstrate efficacy of the μ-selection method.
|Original language||English (US)|
|Number of pages||6|
|Journal||Proceedings of the American Control Conference|
|State||Published - 2004|
All Science Journal Classification (ASJC) codes
- Control and Systems Engineering