Matching of broken random samples with a recurrent neural network

Various problems arising in the reconstruction of high-energy physics events can be solved in the general framework of matching of broken random samples. Typical examples are signal matching, track-track matching, and track-hit matching. If the random samples are complete and free of noise, there exists an optimal statistical procedure. In a real-world application, however, the sample usually is noisy and/or incomplete due to observation inefficiency. In this case the optimal procedure is inapplicable. We investigate a sequential and a global approach to solving the resulting combinatorial problem. The global approach is implemented by a recurrent neural network with mean-field annealing. We analyze the performance of both methods for various types of random samples and various levels of inefficiency and noise.