Merge pull request #49 from PatrickOHara/metricness-londonaq

Fix metricness for londonaq by dividing costs

tests/test_converter/test_split_converter.py

@@ -37,4 +37,4 @@ def test_split_graph_from_properties():
     G = split_graph_from_properties(properties)
     for _, _, data in G.edges(data=True):
         old_edge = data["old_edge"]
-        assert data["cost"] == float(properties[old_edge]["cost"]) / 2.0
+        assert data["cost"] == float(properties[old_edge]["cost"]) or data["cost"] == 0

tests/test_metric.py

@@ -3,6 +3,7 @@
 import networkx as nx
 import pytest
 from tspwplib import build_path_to_oplib_instance, metricness, mst_cost, ProfitsProblem
+from tspwplib.metric import semi_mst_cost
 
 
 def test_mst_cost(oplib_root, generation, graph_name):
@@ -24,6 +25,18 @@ def test_mst_cost(oplib_root, generation, graph_name):
             )
 
 
+def test_semi_mst_cost(oplib_root, generation, graph_name):
+    """Test Semi MST cost"""
+    filepath = build_path_to_oplib_instance(oplib_root, generation, graph_name)
+    problem = ProfitsProblem.load(filepath)
+    G = problem.get_graph()
+    new_cost = semi_mst_cost(G, cost_attr="cost")
+    T = nx.minimum_spanning_tree(G, weight="cost")
+    tree_cost = sum(nx.get_edge_attributes(T, "cost").values())
+    upper_bound_cost = sum(mst_cost(G).values())
+    assert tree_cost < sum(new_cost.values()) < upper_bound_cost
+
+
 @pytest.mark.parametrize(
     "edges,expected_metricness",
     [

tspwplib/converter.py

@@ -419,7 +419,7 @@ def prize_from_weighted_edges(
 
 def split_edge_cost(
     edge_cost: EdgeFunction, to_split: LookupToSplit
-) -> Dict[Edge, float]:
+) -> Dict[Edge, int]:
     """Assign half the cost of the original edge to each of the split edges.
 
     Args:
@@ -436,9 +436,8 @@ def split_edge_cost(
     split_cost = {}
     for edge, cost in edge_cost.items():
         first_split, second_split = to_split[edge]
-        half_cost = float(cost) / 2.0
-        split_cost[first_split] = half_cost
-        split_cost[second_split] = half_cost
+        split_cost[first_split] = cost
+        split_cost[second_split] = 0
     return split_cost
 
 

tspwplib/metric.py

@@ -2,6 +2,7 @@
 
 import random
 import networkx as nx
+import numpy as np
 from .exception import NoTreesException, NotConnectedException
 from .types import SimpleEdgeList, SimpleEdgeFunction
 
@@ -93,3 +94,49 @@ def mst_cost(G: nx.Graph, cost_attr: str = "cost") -> SimpleEdgeFunction:
         else:
             new_cost[(u, v)] = cost + tree_cost[u][v]
     return new_cost
+
+
+def semi_mst_cost(
+    G: nx.Graph, cost_attr: str = "cost", seed: int = 0
+) -> SimpleEdgeFunction:
+    """Half of the non-MST-tree edges are left unchanged.
+    The other half are assigned cost as described in mst_cost.
+
+    The half of edges are chosen with uniform and independent probability.
+
+    Args:
+        G: Undirected, simple graph
+        cost_attr: Name of the cost attribute of edges
+        seed: Set the seed of the random number generator
+
+    Returns
+        A new cost function
+    """
+    # find the cost of the minimum spanning tree in G
+    T = nx.minimum_spanning_tree(G, weight=cost_attr)
+    tree_cost = dict(nx.all_pairs_bellman_ford_path_length(T, weight=cost_attr))
+
+    # get edges not in the tree
+    non_tree_edges = [(u, v) for u, v in G.edges() if not T.has_edge(u, v)]
+    num_non_tree_edges = len(non_tree_edges)
+
+    gen = np.random.default_rng(seed=seed)
+    donot_change_edge_cost = set()
+    while len(non_tree_edges) >= float(num_non_tree_edges) * 0.5:
+        index = gen.integers(0, len(non_tree_edges) - 1)  # get random index
+        donot_change_edge_cost.add(
+            non_tree_edges.pop(index)
+        )  # remove item at random index, add to set
+
+    # set the cost of the new edges
+    new_cost: SimpleEdgeFunction = {}
+    for (u, v), cost in nx.get_edge_attributes(G, cost_attr).items():
+        if (
+            T.has_edge(u, v)
+            or (u, v) in donot_change_edge_cost
+            or (v, u) in donot_change_edge_cost
+        ):
+            new_cost[(u, v)] = cost
+        else:
+            new_cost[(u, v)] = cost + tree_cost[u][v]
+    return new_cost