Hybrid Planning Heuristics Based on Task Decomposition Graphs

Abstract

Hybrid Planning combines Hierarchical Task Network (HTN) planning with concepts known from Partial-Order Causal-Link (POCL) planning. We introduce novel heuristics for Hybrid Planning that estimate the number of necessary modifications to turn a partial plan into a solution. These estimates are based on the task decomposition graph that contains all decompositions of the abstract tasks in the planning domain. Our empirical evaluation shows that the proposed heuristics can significantly improve planning performance.

Tight Bounds for HTN Planning with Task Insertion (Extended Abstract)

“…In our proofs, the acyclic progression technique was instrumental in proving upper bounds that matched the lower bounds for each of the problem classes we examined. As such, it is at least a baseline algorithm to compare against in future work, likely in the context of modern hybrid planners such as PANDA (Bercher, Keen, and Biundo 2014), or the more loosely hierarchical ANML (Smith, Frank, and Cushing 2008).…”

Hierarchical planning and reasoning about partially ordered plans—From theory to practice

“…Early work on domain‐independent heuristics for HTN planning centered around the computation of landmarks (Elkawkagy et al. 2012; Bercher, Keen, and Biundo 2014). This work focused on constructing a TDG (as the one provided in Figure 3) and computed landmarks via taking the intersection of all decomposition methods that belong to the same compound task.…”

“…This task was modeled as a planning problem and solved by our hybrid planner that combines HTN planning with POCL planning (Bercher, Keen, and Biundo 2014; Bercher et al. 2017).…”

Applying Monte-Carlo Tree Search in HTN Planning

“…Traditionally, HTN planning systems relied mostly on the informativeness of the domain knowledge provided, making do with simple search algorithms like depth-first search (Nau et al 2003). Recent work lines have established more informed search mechanisms, namely sophisticated reachability analysis (Bit-Monnot, Smith, and Do 2016; Behnke et al 2020), heuristic search (Bercher, Keen, and Biundo 2014;Bercher et al 2017;Höller et al 2018), compilations to classical planning (Alford, Kuter, and Nau 2009;Alford et al 2016), and SAT (Behnke, Höller, and Biundo 2018;2019a;Schreiber et al 2019). These methods improve HTN planning performance, make it less dependent on the domain knowledge, and provide it with the ability to search for low-cost plans (sometimes with optimality guarantees depending on the technique used, see e.g.…”

“…Our contribution here is the implementation and evaluation of MCTS in HTN planning. We do so in two HTN planning systems, namely Panda (Bercher, Keen, and Biundo 2014) and SHOP (Nau et al 1999).…”