Merge branch 'master' of https://github.com/dourouc05/CombinatorialBa…

…ndits.jl into master

Project.toml

@@ -18,7 +18,7 @@ DataStructures = "^0.16, ^0.17, 0.18"
 Distributions = "^0.16, ^0.17, ^0.18, ^0.19, ^0.20, ^0.21, ^0.22, 0.23, 0.24"
 Hungarian = "^0.6"
 IterTools = "^1.0"
-JuMP = "^0.21"
+JuMP = "^0.21, ^0.22"
 LightGraphs = "^1.0"
 Munkres = "^0.2"
 julia = "^1"

src/policies/ossb.jl

@@ -0,0 +1,43 @@
+# OSSB (optimal sampling for structured bandits).
+# Based on https://arxiv.org/abs/1711.00400.
+
+abstract type OSSBOptimisationAlgorithm end
+function optimise_ossb(instance::CombinatorialInstance{T}, ::OSSBOptimisationAlgorithm,
+                       weights::Dict{T, Float64}) where T end
+
+mutable struct OSSB <: Policy
+  algo::OSSBOptimisationAlgorithm
+end
+
+mutable struct OSSBDetails <: PolicyDetails
+  n_iterations::Int
+  solver_time::Float64
+
+  function OSSBDetails()
+    return new(0, 0.0)
+  end
+end
+
+function choose_action(instance::CombinatorialInstance{T}, policy::ESCB2, state::State{T}; with_trace::Bool=false) where T
+  # For initialisation, if some arms have never been tried, force them to be tried.
+  # This should not be required for OSSB, but it improves running times for the first few iterations.
+  if any(v == 0 for v in values(state.arm_counts))
+    weights = Dict(arm => iszero(state.arm_counts[arm]) ? 1.0 : 0.0 for arm in keys(state.arm_counts))
+
+    t0 = time_ns()
+    sol = solve_linear(instance, weights)
+    t1 = time_ns()
+
+    if with_trace
+      run_details = OSSBDetails()
+      run_details.n_iterations = 1
+      run_details.solver_time = (t1 - t0) / 1_000_000_000
+      return sol, run_details
+    else
+      return sol
+    end
+  end
+
+  # Solve OSSB's problem.
+  return optimise_ossb(instance, policy.algo, state.arm_average_reward, with_trace=with_trace)
+end