Learning the filling policy of a biodegradation process by fuzzy actor-critic learning methodology

Reinforcement Learning is a learning methodology through an agent interacting with the environment. Actor-Critic methods have a separated memory structure to present the independence between the policy and the value function. In systems where the states are defined as continuous, there is a problem with dimensionality, and an approximation method has to be used. The classic Reinforcement Learning algorithms can be combined with Fuzzy Logic techniques to store all value functions, since Fuzzy Logic has been proved to be an effective universal approximator. This work propose a Fuzzy Actor-Critic method to compute and store the state values using fuzzy logic to get a state approximation. Phenol is one of the most important water pollutants on chemistry industries. A phenol biodegradation process consist on a Sequence Batch Reactor (SBR), that need an near optimal filling policy for its correct operation. Fuzzy Actor Critic learning strategy offers a operation policy and it can be linguistically interpreted by the process experts, this approach can be useful to propose a comprehensive filling policy of a biodegradation SBR process. © 2008 Springer Berlin Heidelberg.