Generating decision rules by reinforcement learning for a class of crop management problems
Francisco Ambrosio Garcia, Roger Martin‐Clouaire · 2001
This paper addresses the questions of generating near optimal management strategies for a class of crop management problems. Two approaches are presented. They provide strategies under the form of a timely-structured set of decision rules. Both rely on a formalization of the problem as a finite-horizon Markov decision process and use stochastic optimization techniques that belong to the reinforcement leaming family of algorithms. The basic idea consists in combining the dynamic programming principle with an iterative evaluation mechanism that exploits a simulator of the crop production system. The two approaches enable to deal efficiently with the large size and continuous nature of the state and decision spaces. The first method uses a CMAC discretization methods of these spaces to represent compactly the worth of state-decision pairs and ultimately extracts the decision rules constituting the strategy. The second method generates directly decision rules having fuzzy (flexible) antecedents.