An innovative scheme is proposed for the dynamic control of phase in laser beam array. It is based on a simple neural network that predicts the complex field array from the intensity of the induced scattered pattern through a phase intensity transformer made of a diffuser. A crucial feature of this scheme is the use of a kind of reinforcement learning approach for the NN training which takes account of the iterated corrections. Experiments on a proof of concept system demonstrated the high performance and scalability of the scheme with an array of up to 100 beams.