ELG

the proposed Ensemble of Local and Global policies (ELG) method is a neural solver for Vehicle Routing Problems that integrates a primary global policy, which learns from the complete VRP instance, with an auxiliary local policy that learns from transferable local topological features to boost generalization performance.

ELGInitialization

Bases: Initialization

Methods: ELGInitialization

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

Usage

python train.py setting=elg_settings mode=train problem={problem} local_enable={local_enable} local_size={local_size} xi={xi} euclidean={euclidean}
python train.py settings=elg_settings mode=train problem=tsp local_enable=True local_size=30 xi=-1 euclidean=False
python train.py settings=elg_settings mode=train problem=cvrp local_enable=True local_size=40 xi=-1 euclidean=False

problem	local_enable	local_size	xi	euclidean
tsp	True	30	-1	False
cvrp	True	40	-1	False

Task

Supported Tasks: TSP, CVRP.

Required Data Generator:

• TSPGenerator

• CVRPGenerator

Policy

• class ELGPolicy

class ELGPolicy(nn.Module):
    def __init__(self,
                 env_name: str = "tsp",
                 embed_dim: int = 128,
                 num_heads: int = 8,
                 qkv_dim: int = 16,
                 num_encoder_layers: int = 6,
                 normalization: str = "instance",
                 feedforward_hidden: int = 512,
                 logit_clipping: float = 10,
                 use_graph_mean: bool = False,
                 am_mode: bool = True,
                 first_placeholder: bool = True,
                 local_enable:bool = False,
                 local_size: int = 30,
                 xi: int = -1,
                 euclidean: bool = False,
                 )

• Parameters of ELG

  • general parameter same as POMO

  • local_enable: Enable the local policy

  • local_size: Number of node considered by the local policy

  • xi: Coefficient of distance penalty

  • eucliedean: Whether normalize the distance in the neighborhood

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

Training

ELG applies autoregressive reinforcement learning [calss ARREINFORCELightning]