AI Planning Reformulation for Diverse Non-Redundant Plans
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automated planning systems face challenges in efficiently generating diverse and high-quality plans in complex environments, often resulting in infinite or excessively large solution sets due to indifference in action ordering and unnecessary actions.
Innovation Solution
The approach involves reformulating the planning problem to forbid plans that are supersets of previously found plans, using cost-optimal and agile planners to iteratively generate a set of top-quality or diverse plans by extending equivalence criteria to super-(multi)sets, and utilizing a novel reformulation that forbids plans as super-sets or super-multisets based on operator multisets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional planning algorithms are used to generate all possible plans, then the completeness of solution coverage is improved, but the quantity of plans becomes excessively large or infinite due to indifference in action ordering and unnecessary actions
Solution Approach 1:
The patent changes the equivalence criterion parameter from traditional action sequence equality to super-multiset equality of operators. This parameter change allows the system to group plans that achieve the same goal through different orderings or with different numbers of redundant actions into a single equivalence class, thereby generating a finite set of representative plans while maintaining solution coverage.
Solution Approach 2:
The patent performs preliminary action by defining equivalence classes and selecting representative plans before the actual plan generation process. By pre-establishing the criterion that plans are equivalent if their operator multisets are super-sets of each other, the system avoids generating duplicate or redundant plans during the search process, thus controlling the quantity of output plans.
2Ease of operation
If traditional planning methods are used without reformulation, then the simplicity of the planning process is maintained, but the diversity of generated plans is reduced due to redundant operators and actions
Solution Approach 1:
The patent modifies the equivalence parameter to super-multiset relationship, which fundamentally changes how plan diversity is measured. This parameter change enables the system to recognize and eliminate only truly redundant plans while preserving diverse solution strategies, thereby improving plan diversity without significantly complicating the planning process.
Solution Approach 2:
The patent uses the concept of copying plans with modified equivalence criteria. Instead of generating completely new plans, the system creates representative copies from equivalence classes of plans that are super-sets of each other. This approach maintains simplicity by reusing plan structures while improving diversity through selective representation.
3Reliability
If plans are generated without forbidding super-sets, then the comprehensiveness of solution space exploration is improved, but the efficiency of plan generation deteriorates due to redundant computation
Solution Approach 1:
The patent implements feedback by checking whether newly generated plans are super-sets of previously found plans. This feedback mechanism allows the system to prune the search space dynamically, avoiding redundant computation of plans that have already been represented by shorter or equivalent plans, thus improving generation efficiency while maintaining solution comprehensiveness.
Solution Approach 2:
The patent extracts and removes redundant plans from the solution set by identifying and eliminating plans that are super-sets of previously found plans. This extraction process maintains the comprehensiveness of the solution space by preserving at least one representative from each equivalence class while removing redundant computations, thereby improving efficiency.
Data Source
AI summary
In an approach for improved artificial intelligence planning in an automated machine learning pipeline, a processor formulates an artificial intelligence planning problem. A processor receives a pre-defined stopping criterion for generating one or more plans for the artificial intelligence planning problem. A processor generates the one or more plans by executing a planning algorithm. A processor reformulates the artificial intelligence planning problem into a new artificial intelligence planning problem by forbidding plans that correspond to super-sets of the one or more plans. A processor generates one or more new plans based on the reformulation until the pre-defined stopping criterion is reached.