Combinatorial optimization

Attention-Based Policy Learning for the TSP

Transformer and reinforcement-learning policies for TSP, evaluated for solution quality, generalization, and exploration strategies.

Year
2025
Status
Project
Topics
Transformers · Combinatorial optimization · Reinforcement learning
Multiple candidate tours through the same set of cities

A conceptual illustration of the mathematical idea, not a plot of measured experimental results.

01

The question

How well can an attention-based policy learn to construct short traveling-salesperson tours?

02

Central insight

A learned policy must make a sequence of decisions whose value depends on the completed tour. Evaluating that policy involves both the quality of its solutions and how it behaves on new instances.

03

Approach

  1. 01

    Train attention-based policies for the traveling-salesperson problem.

  2. 02

    Evaluate the quality of the resulting tours and generalization to new instances.

  3. 03

    Compare exploration strategies during policy learning.

04

My contribution

  • Trained transformer/RL policies for TSP.
  • Evaluated solution quality, generalization, and exploration strategies.
05

Result

An experimental study of learned combinatorial decision-making, covering training, solution evaluation, and exploration.

06

What remains

A possible extension is to combine learned tour proposals with classical local search.