.png){kind=icon}
Return
Process Pilot: Learning decision policies from event data for process optimization
.png){kind=icon}
PRE
AI
DOI：10.1016/j.is.2026.102772.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Prescriptive process monitoring (PresPM) studies techniques that leverage event logs to recommend actions at runtime that improve process outcomes. Recent PresPM approaches increasingly rely on reinforcement learning (RL). Because interacting with real-world processes is often infeasible, these approaches typically construct a Markov Decision Process (MDP) from historical event data to simulate the execution environment. However, how to systematically discover such MDPs from event logs and which characteristics define effective RL-based PresPM pipelines remain open problems that limit the reliable deployment of these approaches in practice. This paper presents Process Pilot, an RL-based PresPM approach that discovers an MDP from an event log describing the stochastic environment in which historical processes were executed and uses it to learn a policy for executing future processes in a way that optimizes their outcomes. The paper further proposes two transition-based state abstraction functions for discovered MDPs and introduces measures of the quality of these models. An evaluation using synthetic and industrial event logs compares Process Pilot with prior work. The results show that different abstractions lead to substantially different policy effectiveness, with transition-based abstractions generally outperforming alternatives. Moreover, higher-quality MDP representations often yield recommendations that lead to improved process outcomes.
Keywords:
Prescriptive process monitoring
Reinforcement process learning
Markov decision process discovery and abstraction
Journal
I
IF:
3.4
Papers:
133
Citations:
0
Organization
Cited Papers
No cited papers available