Core Concepts
Cognitive map learners (CMLs) can be assembled into modular, hierarchical systems using hyperdimensional computing (HDC) to solve complex problems like the Tower of Hanoi puzzle without retraining the individual CMLs or explicit knowledge of their graph topologies.
Abstract
The content describes a method for assembling modular, hierarchical cognitive map learners (CMLs) using hyperdimensional computing (HDC) to solve the Tower of Hanoi (ToH) puzzle.
Key highlights:
CMLs are a collection of single-layer neural networks that learn internal representations of node states, edge actions, and edge action availabilities in an abstract graph.
CMLs can perform near-optimal path planning between any two graph node states, but do not learn when or why to transition from one node to another.
This work integrates CMLs with HDC, a form of symbolic machine learning, to create semantically meaningful node state representations that can be shared and orchestrated across multiple CMLs.
Several CML modules were prepared independently and then repurposed to solve the ToH puzzle without retraining the CMLs and without explicit reference to their respective graph topologies.
Four methods were demonstrated for orchestrating the ring CMLs to solve the ToH puzzle: 1) an external, monolithic policy; 2) local, partial policies; 3) mapping ToH states to ring states; and 4) a composite ToH CML constructed from the ring CML states.
This work suggests a template for building levels of biologically plausible cognitive abstraction and orchestration using modular, hierarchical CML-HDC systems.