Saeed Amizadeh - Control and Intelligent Processing Center of Excellence (CIPCE), School of Electrical and Computer Engineering, University of Tehran, Tehran, Iran
Majid Nili Ahmadabadi - Control and Intelligent Processing Center of Excellence (CIPCE), School of Electrical and Computer Engineering, University of Tehran, Tehran, Iran، School of Cognitive Sciences,Institute for studies in theoretical Physics and Mathematics, Tehran, Iran
Caro Lucas - Control and Intelligent Processing Center of Excellence (CIPCE), School of Electrical and Computer Engineering, University of Tehran, Tehran, Iran، School of Cognitive Sciences,Institute for studies in theoretical Physics and Mathematics, Tehran, Iran

Continuous-State Reinforcement Learning (RL) has been recently favored because of the continuous nature of the real world RL problems and many theoretical approaches have been devised to handle the case. However, most of these methods presume that the structure of the agent's perceptual environment is fed to it. But this is not the case in many real situations. Inspired from the subjective view existing in the Cognitive Constructivist learning theory, in this paper, a new method is presented to discover and construct the structure of the environment in parallel with learning the optimal policy. To achieve these goals, the proposed approach incorporates the Bayesian formalism to or ganize the perceptual space while it tries to learn the optimal behavior using a Q-learning-like learning algorithm. These characteristics as a whole define a Reinforcement Learning algorithm which is developed based on a mixture of Cognitive Constructivism and traditional Behaviorism ideas. Simulation results demonstrate the viability and efficiency of the proposed algorithm on continuous state RL problems.

Bayesian Framework, Continuous-State Reinforcement Learning, Cognitive Constructivism, Mixture Densities