ENHSP
ENHSP, the Expressive Numeric Heuristic Search Planner, is a forward search planner for classical, numeric and PDDL+ problems, including non-linear ones. The organisers of the IPC 2023 numeric track built their comparison baselines on it.
- Authors
- Enrico Scala, with Patrik Haslum, Sylvie Thiébaux and Miquel Ramirez on its main papers
- Institutions
- At publication (2020): University of Brescia, Australian National University, University of Melbourne
- Kind of planning
- Classical, numeric, PDDL+ (time discretised)
- Source
- ENHSP homepage, github.com/hstairs/enhsp (GPL-3.0)
How it works
ENHSP searches forward through states, guided by heuristics designed for numeric problems. Its 2016 ECAI paper generalises interval-based relaxation to problems "with non-linear conditions and effects, and cyclic dependencies". Its 2020 JAIR paper decomposes conditions into atomic subgoals treated in isolation, which gives heuristics for both satisficing and optimal planning.
For PDDL+ it discretises time, by default in steps of one second, so a plan is valid for that discretisation. ENHSP is written in Java.
What it handles
Numeric conditions and effects, plan metrics, conditional effects and action costs, with some forms of disjunctive and quantified conditions. It handles PDDL+ processes and events, which Autoplan does not yet, and some non-linear continuous change. Its README states that it does not support durative actions. See the feature comparison.
In the competitions
We could not confirm a competition entry of ENHSP itself. At IPC 2023, the numeric track's baselines were "implemented in the ENHSP planning system", and an ENHSP configuration also appears as a baseline in the IPC 2026 numeric results.
Results in our benchmarks
| Track | Problems | ENHSP | Autoplan |
|---|---|---|---|
| IPC 1998–2018 classical | 2,742 | 1,041 | 2,450 |
| IPC 2023 classical | 140 | 14 | 109 |
| IPC 2023 numeric | 400 | 195 | 383 |
| IPC 2026 numeric | 320 | 147 | 270 |
ENHSP 20-0.21.0: problems solved with a plan the VAL validator accepts, within 60 seconds and 4 GB of memory on one core. These are our own runs, not competition results, and a planner written for other goals or for a different kind of problem can score low here while being the right tool for its purpose. How we measured.
Papers
- Enrico Scala, Patrik Haslum, Sylvie Thiébaux, Miquel Ramirez. Interval-Based Relaxation for General Numeric Planning. ECAI 2016, Frontiers in Artificial Intelligence and Applications 285, 655-663.
- Enrico Scala, Patrik Haslum, Sylvie Thiébaux. Heuristics for Numeric Planning via Subgoaling. IJCAI 2016, 3228-3233.
- Enrico Scala, Patrik Haslum, Sylvie Thiébaux, Miquel Ramirez. Subgoaling Techniques for Satisficing and Optimal Numeric Planning. Journal of Artificial Intelligence Research 68, 691-752, 2020.