Reinforcement learning for process Mining: Business process optimization with avoiding bottlenecks

Process mining extracts knowledge from event data to understand, analyze, and improve processes. The activity of manually identifying the optimal sequence of transitions for process mining is challenging due to several difficulties, including the high complexity of business processes that may involve numerous activities and decision points, and the significant effort required to collect and analyze the necessary data, ex1plore different transition possibilities, and evaluate their impact on process performance. In our study, we have implemented Process Mining Environment in OpenAI gym format aimed at enhancing the development of reinforcement learning algorithms for process optimization tasks. The capabilities of these approaches of reinforcement learning using the Q-learning and Deep Q-network (DQN) techniques to identify the optimal path. This is achieved by constructing a reward matrix tailored to each method, designed to circumvent absorption states that signify bottlenecks. The environment was tested using a proprietary dataset containing 3,414 tickets, with event logs sourced from the ServiceNow ticketing system. The findings indicate a significant reduction in the action space, with Q-learning and DQN achieving a decrease of 75% and 67%, respectively.