The Fast Downward Planning System (1109.6051v1)

Abstract: Fast Downward is a classical planning system based on heuristic search. It can deal with general deterministic planning problems encoded in the propositional fragment of PDDL2.2, including advanced features like ADL conditions and effects and derived predicates (axioms). Like other well-known planners such as HSP and FF, Fast Downward is a progression planner, searching the space of world states of a planning task in the forward direction. However, unlike other PDDL planning systems, Fast Downward does not use the propositional PDDL representation of a planning task directly. Instead, the input is first translated into an alternative representation called multi-valued planning tasks, which makes many of the implicit constraints of a propositional planning task explicit. Exploiting this alternative representation, Fast Downward uses hierarchical decompositions of planning tasks for computing its heuristic function, called the causal graph heuristic, which is very different from traditional HSP-like heuristics based on ignoring negative interactions of operators. In this article, we give a full account of Fast Downwards approach to solving multi-valued planning tasks. We extend our earlier discussion of the causal graph heuristic to tasks involving axioms and conditional effects and present some novel techniques for search control that are used within Fast Downwards best-first search algorithm: preferred operators transfer the idea of helpful actions from local search to global best-first search, deferred evaluation of heuristic functions mitigates the negative effect of large branching factors on search performance, and multi-heuristic best-first search combines several heuristic evaluation functions within a single search algorithm in an orthogonal way. We also describe efficient data structures for fast state expansion (successor generators and axiom evaluators) and present a new non-heuristic search algorithm called focused iterative-broadening search, which utilizes the information encoded in causal graphs in a novel way. Fast Downward has proven remarkably successful: It won the "classical (i.e., propositional, non-optimising) track of the 4th International Planning Competition at ICAPS 2004, following in the footsteps of planners such as FF and LPG. Our experiments show that it also performs very well on the benchmarks of the earlier planning competitions and provide some insights about the usefulness of the new search enhancements.

• The paper introduces a novel conversion from propositional to multi-valued planning tasks that explicitly reveals implicit constraints to enhance efficiency.
• The paper presents a causal graph heuristic and incorporates advanced search enhancements like preferred operators and deferred heuristic evaluation to optimize planning.
• The paper demonstrates competitive performance across 1442 tasks by leveraging multi-heuristic best-first search and specialized data structures for state expansion.

Insights into the Fast Downward Planning System

Fast Downward (FD) is a comprehensive planning system revolving around heuristic search, designed to address general deterministic planning problems using advanced PDDL2.2 features. Authored by Malte Helmert, this system diverges from conventional propositional PDDL representation, opting instead for multi-valued planning tasks, facilitating the explicit articulation of implicit constraints.

Key Contributions

Fast Downward distinguishes itself through several unique innovations:

1. Alternative Representation: Rather than directly utilizing propositional PDDL representation, input tasks are converted into multi-valued planning tasks. This transformation unveils many implicit constraints of traditional propositional planning tasks explicitly, enhancing computational efficiency.
2. Causal Graph Heuristic: FD introduces a causal graph heuristic, diverging from traditional HSP-like heuristics. Instead of ignoring negative interactions of operators, it employs hierarchical decompositions of planning tasks. This approach ensures a more nuanced computation of heuristic functions, addressing subproblems defined by state variables within their causal predecessors.
3. Search Techniques: Several search enhancements are incorporated within Fast Downward's best-first search algorithm:
   • Preferred Operators: This concept extends the idea of helpful actions from local search to global search, prioritizing more promising paths.
   • Deferred Heuristic Evaluation: Strategically postponing the computation of heuristic evaluations mitigates the adverse effects of large branching factors, thereby improving search performance.
   • Multi-Heuristic Best-First Search: Combining several heuristic evaluation functions orthogonally within a single search algorithm provides robustness and adaptability.
   • Focused Iterative-Broadening Search: This novel non-heuristic algorithm utilizes causal graph information to iteratively broaden the search, focusing on the most promising operators.
4. Data Structures for Efficiency: The planning system employs specialized data structures like successor generators and axiom evaluators to achieve computational efficiency in state expansion and derived variable evaluation, respectively.

Empirical Performance

Evaluations were conducted on a benchmark suite comprising of 1442 tasks from previous International Planning Competitions. The results demonstrate that Fast Downward is competitive, often surpassing state-of-the-art planners like FF and LPG in various domains. The highlights include:

• Solving Diverse Planning Tasks: FD successfully tackles STRIPS domains, ADL domains, and more advanced PDDL2.2 domains incorporating axioms.
• Optimization Strategies: The use of preferred operators and multi-heuristic approaches significantly improves performance. In particular, the M+P configuration (multi-heuristic best-first search with preferred operators) consistently performs best, solving the highest number of tasks across different domains.
• Efficiency in Large Domains: Deferred heuristic evaluation and focused iterative-broadening search demonstrate their utility, especially in domains with large branching factors or numerous goals.

Implications and Future Directions

The implications of Fast Downward's methodology are manifold:

• Efficient Hierarchical Planning: By leveraging hierarchical problem decomposition through causal graph analysis, FD offers a scalable approach to solving complex planning problems.
• Adapting Heuristic Guidance: The flexible incorporation of multiple heuristic functions enables a more robust search, catering to diverse problem characteristics.

Looking ahead, several avenues for further research are evident:

1. Enhanced Heuristic Development: Development of more sophisticated heuristics that can handle cyclic causal graphs without pruning, improving heuristic accuracy.
2. Dynamic State Representations: Experimenting with different state representations and encodings for optimized domain-specific performance.
3. Comprehensive Goal Ordering Techniques: Incorporating advanced goal ordering techniques to further enhance planning performance.
4. Iterative Search Enhancements: Investigating new search strategies that focus on evaluating operator usefulness locally, potentially minimizing overlooked beneficial paths.

In summary, Fast Downward represents a significant advance in AI planning, offering robust and efficient solutions through its unique problem decomposition and heuristic strategies. Its performance in benchmark evaluations underscores both its theoretical innovation and practical applicability, paving the way for future developments in hierarchical planning and heuristic search methodologies.