Abstract:Pattern database (PDB) is one of the most popular automated heuristic generation techniques. A PDB maps states in a planning task to abstract states by considering a subset of variables and stores their optimal costs to the abstract goal in a look up table. As the result of the progress made on symbolic search over recent years, symbolic-PDB-based planners achieved impressive results in the International Planning Competition (IPC) 2018. Among them, Complementary 1 (CPC1) tied as the second best planners and the best non-portfolio planners in the cost optimal track, only 2 tasks behind the winner. It uses a combination of different pattern generation algorithms to construct PDBs that are complementary to existing ones. As shown in the post contest experiments, there is room for improvement. In this paper, we would like to present our work on refining the PDB construction mechanism of CPC1. By testing on IPC 2018 benchmarks, the results show that a significant improvement is made on our modified planner over the original version.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to improve the construction mechanism of the Complementary Pattern Database (CPC) in order to enhance its performance in cost - optimal planning tasks. Specifically, the paper aims to optimize and improve the PDB construction algorithm and its evaluator used in CPC1, and better integrate all components, thereby significantly improving the performance of the planner.
### Background problems
The Pattern Database (PDB) is a commonly - used automated heuristic generation technique. By considering a set of variables (i.e., patterns) in the task, it maps the original state to an abstract state and stores the optimal cost for these states to reach the abstract goal. PDB performed well in the International Planning Competition (IPC) 2018, especially the Complementary 1 (CPC1) planner, which achieved second place in the cost - optimal track, only two tasks behind the champion. However, post - competition experiments revealed some undesirable behaviors of CPC1. For example, using a single PDB construction algorithm (such as GAMER - Style) can solve more tasks, indicating that these algorithms are not well integrated.
### Main contributions of the paper
In order to improve the PDB construction mechanism of CPC1, the authors propose the following improvements:
1. **Optimizing the PDB construction algorithm**:
- **GAMER - Style**: Changing from the average heuristic value to random - walk sampling, which improves the accuracy of the heuristic.
- **Next - Fit Bin - Packing** and **Causal Dependency Bin - Packing (CBP)**: Introducing more efficient randomization and data structures to generate more diverse PDBs.
2. **Improving the evaluator**:
- Using a random - walk sampling evaluator and resampling according to the change in the initial heuristic value to ensure the accuracy and flexibility of the evaluation.
3. **Adaptive PDB construction process**:
- Introducing the UCB1 formula to balance exploration and exploitation, and select the most suitable PDB construction algorithm and its parameter configuration.
- In each iteration, dynamically adjust the PDB construction process according to the time limit and memory limit, and prune the redundant PDB sets to reduce the time cost of subsequent evaluations.
### Experimental results
Through experiments on the IPC 2018 benchmark, the results show that the improved planner CPC0 is superior to the original version CPC1 in several aspects:
- **Higher initial heuristic value**: CPC0 can usually generate PDBs with higher initial heuristic values, indicating that the improved algorithms and evaluators can generate better heuristics.
- **Fewer node expansions**: Although CPC0 has slightly more node expansions in some tasks, overall, CPC0 has fewer node expansions.
- **Faster search time**: The average search time of CPC0 is significantly reduced, mainly due to better heuristics and faster heuristic look - up.
- **Higher task coverage**: CPC0 solves 12 additional tasks, further proving its competitiveness.
In summary, this paper significantly improves the performance of the planner in cost - optimal planning tasks by optimizing the PDB construction mechanism, and once again verifies the effectiveness of the CPC mechanism.