A learning sequential detection method based on neural networks

This paper presents a neural network method for sequential detection. A theorem is stated that there exists a reinforcement learning algorithm which can approach the performance of the optimal sequential probability ratio test (SPRT) in the minimum mean squared-error. Then a suitable network architecture and learning algorithm are developed to implement the reinforcement learning. Simulations have shown that this learning detector can operate as optimum as SPRT while using much less statistical knowledge than SPRT.