update

src/mcts.jl

@@ -2,7 +2,7 @@ module mcts
   
 export selectBestNextNode, selectBestTrajectoryNode, backpropagate, isleaf, isroot, selectChildNode,
         expand, simulate, makeNewState
-
+using Base.Threads
 using GeneralUtils
 
 using ..type
@@ -196,37 +196,69 @@ end
 # Return
   - None
 
+# TODO
+  - [WORKING] implement multithreads
+
 # Signature
 """
-function expand(node::MCTSNode,transition::Function, transitionargs::NamedTuple; 
+function expand(node::MCTSNode, transition::Function, transitionargs::NamedTuple; 
           totalsample::Integer=3)
-
-  nthSample = 0
-  while true
-    nthSample += 1
-    if nthSample <= totalsample
-      result = transition(node.state, transitionargs)
-      newNodeKey::AbstractString = result[:newNodeKey]
-      newstate::AbstractDict = result[:newstate]
-      progressvalue::Integer = result[:progressvalue]
-
-      """
-      [] newNodeKey ∉ keys(node.children). 
-            New state may have semantic vector close enought to 
-          one of existing child state. Which can be assume that they are the same state 
-          semantically-wise i.e. De javu. This could be used to recall lessons for this 
-          similar situation to improve decisionMaker and evaluator. 
-      """
-      if newNodeKey ∉ keys(node.children)
-        node.children[newNodeKey] = 
-          MCTSNode(newNodeKey, newstate, 0, progressvalue, 0, newstate[:reward], 
-            newstate[:isterminal], node, Dict{String, MCTSNode}())
+  results = Any[]
+  @sync for i in 1:totalsample
+      @spawn begin
+          results[i] = transition(deepcopy(node.state), deepcopy(transitionargs))
       end
-    else
-      break
+      println("--> sampling $i")
+  end
+
+  for result in results
+    newNodeKey::AbstractString = result[:newNodeKey]
+    newstate::AbstractDict = result[:newstate]
+    progressvalue::Integer = result[:progressvalue]
+
+    """
+    [] newNodeKey ∉ keys(node.children). 
+          New state may have semantic vector close enought to 
+        one of existing child state. Which can be assume that they are the same state 
+        semantically-wise i.e. De javu. This could be used to recall lessons for this 
+        similar situation to improve decisionMaker and evaluator. 
+    """
+    if newNodeKey ∉ keys(node.children)
+      node.children[newNodeKey] = 
+        MCTSNode(newNodeKey, newstate, 0, progressvalue, 0, newstate[:reward], 
+          newstate[:isterminal], node, Dict{String, MCTSNode}())
     end
   end
 end
+# function expand(node::MCTSNode,transition::Function, transitionargs::NamedTuple; 
+#           totalsample::Integer=3)
+
+#   nthSample = 0
+#   while true
+#     nthSample += 1
+#     if nthSample <= totalsample
+#       result = transition(node.state, transitionargs)
+#       newNodeKey::AbstractString = result[:newNodeKey]
+#       newstate::AbstractDict = result[:newstate]
+#       progressvalue::Integer = result[:progressvalue]
+
+#       """
+#       [] newNodeKey ∉ keys(node.children). 
+#             New state may have semantic vector close enought to 
+#           one of existing child state. Which can be assume that they are the same state 
+#           semantically-wise i.e. De javu. This could be used to recall lessons for this 
+#           similar situation to improve decisionMaker and evaluator. 
+#       """
+#       if newNodeKey ∉ keys(node.children)
+#         node.children[newNodeKey] = 
+#           MCTSNode(newNodeKey, newstate, 0, progressvalue, 0, newstate[:reward], 
+#             newstate[:isterminal], node, Dict{String, MCTSNode}())
+#       end
+#     else
+#       break
+#     end
+#   end
+# end
 
 
 """ Simulate interactions between agent and environment