Core Concepts
The circuit complexity of goal-conditioned policies for classical planning problems can be characterized by the regression width of the planning problems, which determines the necessary size and depth of relational neural network circuits to represent such policies.
Abstract
The paper analyzes the circuit complexity of goal-conditioned policies for classical planning problems, represented as relational neural networks (RelNNs). It makes the following key contributions:
It introduces the concept of serialized goal regression search (S-GRS) and defines the notion of strong optimally-serializable (SOS) width, which characterizes the complexity of planning problems. Problems with constant SOS width can be solved efficiently using S-GRS.
It shows how to compile the BWD and S-GRS algorithms into RelNN circuits, where the circuit complexity depends on the SOS width of the planning problem. Problems with constant SOS width can be compiled into finite-breadth, possibly unbounded-depth RelNN circuits.
It introduces the concept of a regression rule selector (RRS), which can further reduce the circuit complexity by directly predicting the appropriate regression rule to use. If the RRS can be computed by a finite-depth, finite-breadth RelNN, then the overall policy can be compiled into a shallow RelNN circuit.
It analyzes the SOS width and circuit complexity for several classical planning problems, including Blocks World, Logistics, Gripper, Rover, Elevator, and Sokoban. It shows that most problems have constant SOS width except for Sokoban, which has unbounded width.
It presents experimental results on learning RelNN policies for some of these planning problems, demonstrating that the theoretical predictions about circuit complexity match the empirical performance.
The paper provides a principled understanding of the circuit complexity of goal-conditioned policies for classical planning problems, and suggests insights into why certain planning problems are harder than others.
Stats
There are no key metrics or important figures used to support the author's key logics.
Quotes
There are no striking quotes supporting the author's key logics.