Reinforcement Learning for Solving the Vehicle Routing Problem

This repository contains pytorch implementation for solving VRP using Reinforcement Learning[1]

Reward Function

1. Advantage = total distance of predicted route - critic value
2. Actor = advantage* log probability
3. Critic = advantage^2

Model Details

Actor

Encoder

CNN Embedding

Decoder

Single LSTM Layer

Critic

1. 2 Dense Layers

Optimizer

1. SGD
2. learning rate 1.0
3. L2 Gradient Clipping 2.0
4. Batch Size 128

Hardware

1. OS: macOS 14.0 23A344 arm64
2. Host: MacBookPro17,1
3. CPU: Apple M1
4. GPU: Apple M1
5. Memory: 16384MiB

Links

[1] Nazari, M. et al.: Reinforcement Learning for Solving the Vehicle Routing Problem.