Learning Finite Automata Using Local Distinguishing Experiments

One of the open problems listed in Rivest and Schapire, 1989] is whether and how that the copies of L in their algorithm can be combined into one for better performance. This paper describes an algorithm called D that does that combination. The idea is to represent the states ofthe learned modelusing observable symbolsas well as hidden symbolsthat are constructed during learning. These hidden symbols are created to re ect the distinct behaviors of the model states. The distinct behaviors are represented as local distinguishing experiments (LDEs) (not to be confused with global distinguishing sequences), and these LDEs are created when the learner's prediction mismatches the actual observation from the unknown machine. To synchronize the model with the environment, these LDEs can also be concatenated to form a homing sequence. It can be shown that D can learn, with probability 1 ; , a model that is an -approximation of the unknown machine, in a number of actions polynomial in the ...

• Type: Conference
• Year: 1993
• Where: IJCAI
• Authors: Wei-Mein Shen
• Keywords: Distinct Behaviors, Global Distinguishing Sequences, IJCAI 1993, IJCAI 2007, Unknown Machine