Reinforcement Learning in the Load Balancing Problem for the iFDAQ of the COMPASS Experiment at CERN

Currently, modern experiments in high energy physics impose great demands on the reliability, efficiency, and data rate of Data Acquisition Systems (DAQ). The paper deals with the Load Balancing (LB) problem of the intelligent, FPGA-based Data Acquisition System (iFDAQ) of the COMPASS experiment at CERN and presents a methodology applied in finding optimal solution. Machine learning approaches, seen as a subfield of artificial intelligence, have become crucial for many well-known optimization problems in recent years. Therefore, algorithms based on machine learning are worth investigating with respect to the LB problem. Reinforcement learning (RL) represents a machine learning search technique using an agent interacting with an environment so as to maximize certain notion of cumulative reward. In terms of RL, the LB problem is considered as a multi-stage decision making problem. Thus, the RL proposal consists of a learning algorithm using an adaptive e-greedy strategy and a policy retrieval algorithm building a comprehensive search framework. Finally, the performance of the proposed RL approach is examined on two LB test cases and compared with other LB solution methods.