A PDDL+ Benchmark Problem: The Batch Chemical Plant

Giuseppe Della Penna, Benedetto Intrigila, Daniele Magazzeni, Fabio Mercorio

Abstract

The PDDL+ language has been mainly devised to allow modelling of real-world systems, with continuous, time-dependant dynamics. Several interesting case studies with these characteristics have been also proposed, to test the language expressiveness and the capabilities of the support tools. However, most of these case studies have not been completely developed so far. In this paper we focus on the batch chemical plant case study, a very complex hybrid system with nonlinear dynamics that could represent a challenging benchmark problem for planning techniques and tools. We present a complete PDDL+ model for such system, and show an example application where the UPMurphi universal planner is used to generate a set of production policies for the plant.

Introduction

In the field of automated planning, traditional benchmarks often focus on discrete domains like logistics or block manipulation. However, industrial environments such as chemical plants are inherently hybrid, requiring the coordination of discrete actions (e.g., toggling valves) with continuous, time-dependent physical processes (e.g., fluid flow and thermal changes). While PDDL+ was developed to bridge this gap, the community has lacked high-fidelity, complex benchmarks to rigorously test hybrid planners. This work addresses this by providing a comprehensive PDDL+ model for the "Batch Chemical Plant," a classic problem from the Verification of Hybrid Systems (VHS) project.

Methodology

The modeled plant is a sophisticated system comprising seven tanks, twenty-six valves, two pumps, and various heating and cooling elements.

Schematic diagram of the batch chemical plant benchmark, showing tanks (B1-B7), valves (V1-V29), and pumps (P1-P2).

To capture the plant's dynamics, the authors utilized three primary PDDL+ constructs:

  • Durative Actions: These represent operations like filling tanks, where the planner uses duration inequalities to determine the optimal cutoff point based on state variables like concentration.
  • Processes: These model continuous phenomena, such as the flow of saline solution or evaporation during boiling.
  • Events: These handle instantaneous state changes, particularly safety violations (e.g., tank overflows) which invalidate the current plan.

A technical challenge involved PDDL+'s lack of native support for non-linear functions (e.g., square roots for fluid dynamics). The authors resolved this by implementing a second-degree polynomial approximation to maintain physical accuracy within the language's constraints. The resulting model includes 59 predicates, 55 functions, 19 events, and 11 continuous processes.

Results and Discussion

The model was tested using the UPMurphi universal planner. Unlike traditional planners that output a single sequence of actions, UPMurphi generates a policy—a mapping of optimal actions for any reachable state. The planner successfully navigated a massive state space to find viable production policies. By testing against a "cloud" of initial solution volumes (1.5 to 3.7 liters), the authors demonstrated the robustness of the generated policies.

Table summarizing the state space complexity and plan performance, showing nearly 30 million reachable states and optimal plan lengths.

As shown in the results, the planner identified approximately 30 million reachable states and generated a policy for over 7 million states leading to the goal. The total computation time was approximately 6,000 seconds, resulting in a 246.3 MB universal plan. Production cycles were optimized to last between 1,920 and 2,200 seconds depending on initial conditions.

Conclusion

This work provides the first complete PDDL+ benchmark for a complex batch chemical plant, offering a rigorous testbed for hybrid planning tools. While the current implementation relies on discretization (10-second time steps and single-decimal precision for variables), it effectively demonstrates the feasibility of generating robust policies for non-linear industrial systems. Future work will focus on integrating the startup phase and comparing the performance of other emerging hybrid planners against this benchmark.

Summary updated

Abstract

The PDDL+ language has been mainly devised to allow modelling of real-world systems, with continuous, time-dependant dynamics. Several interesting case studies with these characteristics have been also proposed, to test the language expressiveness and the capabilities of the support tools. However, most of these case studies have not been completely developed so far. In this paper we focus on the batch chemical plant case study, a very complex hybrid system with nonlinear dynamics that could represent a challenging benchmark problem for planning techniques and tools. We present a complete PDDL+ model for such system, and show an example application where the UPMurphi universal planner is used to generate a set of production policies for the plant.
BibTeX
@inproceedings{penna2021pddl,
  author = {Giuseppe Della Penna and Benedetto Intrigila and Daniele Magazzeni and Fabio Mercorio},
  title = {{A PDDL+ Benchmark Problem: The Batch Chemical Plant}},
  booktitle = {Proceedings of the International Conference on Automated Planning and Scheduling},
  year = {2021},
  doi = {10.1609/icaps.v20i1.13430},
}
Figure