A novel learning automata game with local feedback for parallel optimization of hydropower production

Abstract

Hydropower optimization for multi-reservoir systems is classi ed as a combinatorial optimization problem with large state-space that is particularly di cult to solve. There exist no golden standard when solving such problems, and many proposed algorithms are domain speci c. The literature describes several di erent techniques where linear programming approaches are extensively discussed, but tends to succumb to the curse of dimensionality problem when the state vector dimensions increase. This thesis introduces LA LCS, a novel learning automata algorithm that utilizes a parallel form of local feedback. This enables each individual automaton to receive direct feedback, resulting in faster convergence. In addition, the algorithm is implemented using a parallel architecture on a CUDA enabled GPU, along with exhaustive and random search. LA LCS has been veri ed through several scenarios. Experiments show that the algorithm is able to quickly adapt and nd optimal production strategies for problems of variable complexity. The algorithm is empirically veri ed and shown to hold great promise for solving optimization problems, including hydropower production strategies.