Reinforcement Learning for Process Mining: Business Process Optimization

Process mining aims to extract knowledge from event data to understand, analyze, and improve Processes. Utilizing the benefits of Reinforcement Learning enables the automation of the business processes discovery, by systematically exploring the state space based on a certain environment interaction to achieve a certain goal, that reduces time and effort required for manual analysis and decision-making. This paper proposes an automated system capable of discovering certain organization’s business processes and identifying the optimal sequence of transitions towards the ticket terminal state avoiding the bottleneck. To identify the structure of the business processes, the available source/target transitions are extracted, along with the computed transition probabilities between each possible transition. These probabilities are then incorporated into the reward design, which is needed to adjust the agent’s behavior. By applying the Q-Learning algorithm, a Q-function learns the quality (q-value) of taking each possible action, given a certain state, encoded as a table that is updated iteratively during the training until reaching an optimal policy. Finally, after a certain number of episodes, the optimal sequence of transitions is identified. The Deep Q-Network (DQN) algorithm is applied to compare the results between the two approaches, and the Q-Learning agent tends to produce the optimal path that is aligned with the business processes.