The value of learning rate and its change mechanisms is one of the issues in designing learning systems such as learning automata. In most cases a time-based reduction function is used to adjust the learning rate aim at reaching stability in training system. So the lear More
The value of learning rate and its change mechanisms is one of the issues in designing learning systems such as learning automata. In most cases a time-based reduction function is used to adjust the learning rate aim at reaching stability in training system. So the learning rate is a parameter that determines to what extent a learning system is based on past experiences, and the impact of current events on it. This method is efficient but does not properly function in dynamic and non-stationary environments.
In this paper, a new method for adaptive learning rate adjustment in learning automata is proposed. In this method, in addition to the length of time to learn, some statistical characteristics of actions probability vector of Learning Automata are used to determine the increase or decrease of learning rate. Furthermore, unlike existing methods, during the process of learning, both increase and decrease of the learning rate is done and Learning Automata responds effectively to changes in the dynamic random environment.
Empirical studies show that the proposed method has more flexibility in compatibility to the non-stationary dynamic environments and get out of local maximum points and the learned values are closer to the true values.
Manuscript profile
In this paper a new structure of learning automata which is called as extended distributed learning automata (eDLA) is introduced. A new eDLA-based iterative sampling method for finding optimal sub-graph in stochastic graphs is proposed. Some mathematical analysis of th More
In this paper a new structure of learning automata which is called as extended distributed learning automata (eDLA) is introduced. A new eDLA-based iterative sampling method for finding optimal sub-graph in stochastic graphs is proposed. Some mathematical analysis of the proposed algorithm is presented and the convergence property of the algorithm is studied. Our study shows that the proposed algorithm can be converge to the optimal sub-graph.
Manuscript profile
In this paper a new learning automata-based algorithm is proposed for learning of parameters of a Bayesian network. For this purpose, a new team of learning automata which is called eDLA is used. In this paper the structure of Bayesian network is assumed to be fixed. Ne More
In this paper a new learning automata-based algorithm is proposed for learning of parameters of a Bayesian network. For this purpose, a new team of learning automata which is called eDLA is used. In this paper the structure of Bayesian network is assumed to be fixed. New arriving sample plays role of the random environment and the accuracy of the current parameters generates the random environment reinforcement signal. Linear algorithm is used to update the action selection probability of the automata. Another key issue in Bayesian networks is parameter learning under circumstances that new samples are incomplete. It is shown that new proposed method can be used in this situation. The experiments show that the accuracy of the proposed automata based algorithm is the same as the traditional enumerative methods such as EM. In addition to the online learning characteristics, the proposed algorithm is in accordance with the conditions in which the data are incomplete and due to the use of learning automaton, has a little computational overhead.
Manuscript profile