Merge pull request #49 from harshangrjn/heuristics

Heuristics for LaplacianOpt

CHANGELOG.md

@@ -1,6 +1,13 @@
 LaplacianOpt.jl Change Log
 =========================
 
+### v0.6.0
+- Added mutiple heuristics to handle both spanning trees and graphs with loops 
+- Refactored `log.jl` to handle solutions from heuristics
+- Included more user options for heuristic in `model_options`
+- Cleaned up populating and logging of eigen cuts
+- Updated docs and unit tests to reflect above changes
+
 ### v0.5.0
 - Generalized eigen cuts to handle any size of non-negative minors (`2x2`-`NxN`)
 - Dropped support for `constraint_eigen_cuts_on_2minors` and `constraint_eigen_cuts_on_3minors`

Project.toml

@@ -1,7 +1,7 @@
 name = "LaplacianOpt"
 uuid = "bb20392f-64fb-4001-92e8-14b3aedd5a9e"
 authors = ["Harsha Nagarajan"]
-version = "0.5.0"
+version = "0.6.0"
 
 [deps]
 DataStructures = "864edb3b-99cc-5e75-8d2d-829cb0a9cfe8"
@@ -11,9 +11,9 @@ JuMP = "4076af6c-e467-56ae-b986-b466b2749572"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MathOptInterface = "b8f27783-ece8-5eb3-8dc8-9495eed66fee"
 Memento = "f28f55f0-a522-5efc-85c2-fe41dfb9b2d9"
+Pkg = "44cfe95a-1eb2-52ea-b672-e2afdf69b78f"
 TikzGraphs = "b4f28e30-c73f-5eaf-a395-8a9db949a742"
 TikzPictures = "37f6aa50-8035-52d0-81c2-5a1d08754b2d"
-Pkg = "44cfe95a-1eb2-52ea-b672-e2afdf69b78f"
 
 [compat]
 DataStructures = "~0.17, ~0.18"

docs/src/quickguide.md

@@ -48,6 +48,9 @@ results = LOpt.run_LOpt(params, lopt_optimizer)
 !!! tip
     Run times of [LaplacianOpt](https://github.com/harshangrjn/LaplacianOpt.jl)'s mathematical optimization models are significantly faster using [Gurobi](https://www.gurobi.com) as the underlying mixed-integer programming (MIP) solver. Note that this solver's individual-usage license is available [free](https://www.gurobi.com/academia/academic-program-and-licenses/) for academic purposes. 
 
+!!! tip
+    Note that [LaplacianOpt](https://github.com/harshangrjn/LaplacianOpt.jl) tries to find the global solution of a combinatiorial optimization problem, that is known to be [NP-hard to compute](https://doi.org/10.1016/j.orl.2008.09.001) in the size of `num_nodes`. To obtain quick feasible solutions via "k-opt-based" heristics, just set `:solution_type` to `"heuristics"` and adjust the `num_swaps_bound_kopt` parameter to appropriate values, where larger value implies a better-quality feasible solution at the cost of the heuristic run time. 
+
 # Extracting results
 The run commands (for example, `run_LOpt`) in LaplacianOpt return detailed results in the form of a dictionary. This dictionary can be used for further processing of the results. For example, for the given instance of a complete graph, the algorithm's runtime and the optimal objective value (maximum algebraic connectivity) can be accessed with,
 

examples/instances/5_nodes/5_11.json

@@ -0,0 +1,4 @@
+{"edges_to_augment":[[[1,5],5.1749106],[[2,3],6.1501476],[[2,4],13.66149798]],
+"num_nodes":5,
+"edges_existing":[[[1,2],34.99664955],[[1,3],3.86014785],[[1,4],11.2685551],[[2,5],21.90200648],[[3,4],29.11616571],[[3,5],27.74393888],[[4,5],2.83029978]]
+}

examples/run_examples.jl

@@ -1,15 +1,14 @@
 import LaplacianOpt as LOpt
 using JuMP
 using CPLEX
-# using Gurobi
 # using GLPK
 
 include("optimizer.jl")
 
-#----------------------------------#
-#            MIP solver            #
-# (Cplex 22.1 performs the best)   #
-#----------------------------------#
+#-----------------------------------#
+#            MIP solver             #
+# (> Cplex 22.1 performs the best)  #
+#-----------------------------------#
 lopt_optimizer = get_cplex()
 
 #-------------------------------------#
@@ -47,6 +46,7 @@ end
 num_nodes = 8
 instance = 1
 data_dict, augment_budget = data_I(num_nodes, instance)
+# data_dict, augment_budget = data_II()
 
 params = Dict{String,Any}("data_dict" => data_dict, "augment_budget" => augment_budget)
 
@@ -57,7 +57,7 @@ params = Dict{String,Any}("data_dict" => data_dict, "augment_budget" => augment_
 
 # For more model options, check https://github.com/harshangrjn/LaplacianOpt.jl/blob/master/src/types.jl
 model_options = Dict{Symbol,Any}(
-    :eigen_cuts_sizes => [2, num_nodes],
+    :eigen_cuts_sizes => [data_dict["num_nodes"], 2],
     :topology_flow_cuts => true,
     :solution_type => "optimal",
 )
@@ -67,5 +67,5 @@ result = LOpt.run_LOpt(
     lopt_optimizer;
     options = model_options,
     visualize_solution = false,  # Make it true to plot the graph solution
-    visualizing_tool = "tikz",   # "graphviz" is another option
+    visualizing_tool = "tikz",   # "tikz", "graphviz" are the options
 )

src/constraints.jl

@@ -93,8 +93,8 @@ function constraint_lazycallback_wrapper(lom::LaplacianOptModel; optimizer = not
                 "Callback status: unknown - solution may not be integer feasible",
             )
 
-        elseif status in
-               [MOI.CALLBACK_NODE_STATUS_INTEGER, MOI.CALLBACK_NODE_STATUS_FRACTIONAL]
+        elseif status in [MOI.CALLBACK_NODE_STATUS_INTEGER]
+            #    [MOI.CALLBACK_NODE_STATUS_INTEGER, MOI.CALLBACK_NODE_STATUS_FRACTIONAL]
             if !(lom.options.formulation_type == "max_span_tree")
                 if (
                     size(lom.options.eigen_cuts_sizes)[1] > 0 &&
@@ -133,9 +133,8 @@ function constraint_lazycallback_wrapper(lom::LaplacianOptModel; optimizer = not
 end
 
 function constraint_eigen_cuts(W_val::Matrix{<:Number}, cb_cuts, lom::LaplacianOptModel)
-    if (size(lom.options.eigen_cuts_sizes)[1] > 0) &&
-       (minimum(lom.options.eigen_cuts_sizes) >= 2)
-        for cut_size in sort(unique(lom.options.eigen_cuts_sizes), rev = true)
+    if (length(keys(lom.minor_idx_dict)) > 0)
+        for cut_size in sort(collect(keys(lom.minor_idx_dict)), rev = true)
             for i in lom.minor_idx_dict[cut_size]
                 LOpt._add_eigen_cut_lazy(W_val, cb_cuts, lom, collect(i))
             end