Merge branch 'master' of https://github.com/JuliaPOMDP/MCTS.jl

src/dpw.jl

@@ -16,6 +16,12 @@ end
 Call simulate and chooses the approximate best action from the reward approximations
 """
 function POMDPs.action(p::DPWPlanner, s)
+    if isterminal(p.mdp, s)
+        error("""
+              MCTS cannot handle terminal states. action was called with
+              s = $s
+              """)
+    end
     S = state_type(p.mdp)
     A = action_type(p.mdp)
     if p.solver.keep_tree
@@ -46,7 +52,7 @@ function POMDPs.action(p::DPWPlanner, s)
         end
     end
     # XXX some publications say to choose action that has been visited the most
-    return tree.a_labels[sanode] # choose action with highest approximate value 
+    return tree.a_labels[sanode] # choose action with highest approximate value
 end
 
 

test/runtests.jl

@@ -86,4 +86,17 @@ let
     @test abs(t-1.0) < 0.5
 end
 
+# test terminal state error
+let
+    solver = DPWSolver(n_iterations=typemax(Int),
+                       depth=depth,
+                       max_time=1.0,
+                       exploration_constant=ec)
+    mdp = GridWorld()
+
+    policy = solve(solver, mdp)
+    state = GridWorldState(1,1,true)
+    @test_throws ErrorException action(policy, state)
+end
+
 nbinclude("../notebooks/Domain_Knowledge_Example.ipynb")

test/visualization.jl

@@ -11,7 +11,7 @@ a = action(policy, state)
 tree = D3Tree(policy, state)
 
 # dpw
-solver = DPWSolver(n_iterations=n_iter, depth=depth, exploration_constant=ec)
+solver = DPWSolver(n_iterations=n_iter, depth=depth, exploration_constant=ec, rng=MersenneTwister(13))
 mdp = GridWorld()
 
 policy = solve(solver, mdp)