INViT

the proposed Invariant Nested View Transformer (INViT) method is a neural solver for Vehicle Routing Problems.

INViTInitialization

Bases: Initialization

Methods: INViTInitialization

The inputs consist of a tensor containing (x, y) coordinates of the instance. INViT then construct a routing solution step by step. After INViTInitializaiton, it outputs a solution solved by the model.

Usage

python train.py setting=invit_settings mode=test problem={problem}

Task

Supported Tasks: TSP, CVRP.

Required Data Generator:

• TSPGenerator

• CVRPGenerator

Policy

• class INViTPolicy

class INVITPolicy(nn.Module):
    def __init__(
        self,
        # node_dim: int = 2,
        env_name: str = "tsp",
        embed_dim: int = 128,
        num_heads: int = 8,
        qkv_dim: int = 16,
        feedforward_dim: int = 512,
        action_embed_layer_num: int = 2,
        state_embed_layer_num: int = 2,
        decoder_layer_num: int = 3,
        normalization: str = "layer",
        first_placeholder: bool = False,
        state_k: list = [35, 50, 65], # directly from the INViT code
        action_k: int = 15,
        logit_clipping: float = 10
    ):

• Parameters of INViT

  • general parameter same as POMO

  • action_embed_layer_num: number of layer of action embedding.

  • state_embed_layer_num: number of layer of state embedding.

  • state_k: size of K-Nearest-Neighbor (KNN)

• Backbone: INViT utilize the Transformer to predict next node.

Training

INViT applies autoregressive reinforcement learning [calss ARREINFORCELightning]