CUCB: allow tuning the confidence radius.

src/policies/cucb.jl

@@ -1,20 +1,24 @@
 # CUCB, "combinatorial upper confidence bound".
 # Based on http://proceedings.mlr.press/v28/chen13a.html.
 
-mutable struct CUCB <: Policy end
+mutable struct CUCB <: Policy 
+  α::Float64
+end
+
+CUCB() = CUCB(sqrt(0.5))
 
 struct CUCBDetails <: PolicyDetails
   solver_time::Float64
 end
 
-function choose_action(instance::CombinatorialInstance{T}, ::CUCB, state::State{T}; with_trace::Bool=false) where T
+function choose_action(instance::CombinatorialInstance{T}, algo::CUCB, state::State{T}; with_trace::Bool=false) where T
   if any(v == 0 for v in values(state.arm_counts))
     # There is at least one arm that has never been tried: maximise the arms that have never been tested.
     weights = Dict(arm => (state.arm_counts[arm] == 0.0) ? 1.0 : 0.0 for arm in keys(state.arm_counts))
   else
     # All arms have been seen, thus this formula makes sense (no zero-valued arm count).
-    weights = Dict(arm => state.arm_average_reward[arm] + sqrt((log(state.round)) / (2 * state.arm_counts[arm])) for arm in keys(state.arm_counts))
-    # weights = Dict(arm => state.arm_average_reward[arm] + sqrt((3 * log(state.round)) / (2 * state.arm_counts[arm])) for arm in keys(state.arm_counts))
+    numerator = algo.α * sqrt(log(state.round))
+    weights = Dict(arm => state.arm_average_reward[arm] + numerator * sqrt(1.0 / state.arm_counts[arm]) for arm in keys(state.arm_counts))
   end
 
   t0 = time_ns()