update

src/interface.jl

@@ -97,6 +97,18 @@ function decisionMaker(a::T1, state::T2)::Dict{Symbol, Any} where {T1<:agent, T2
   # (trajectories)
   # """
 
+
+  """
+  {
+    "Question": "I would like to buy a sedan.",
+    "Thought_1": "I have many cars in my inventory suitable for several usage scenarios.",
+    "Thought_2": "It would be better if I knew what the user intends to do with his car.",
+    "Thought_3": "I will ask the user what is the intended usecase",
+    "Action_1": {"name": "chatbox", "input": "What will you use it for?"}
+  }
+  """
+
+
   _prompt = 
   """
   You are a helpful sommelier working for a wine store. 
@@ -131,19 +143,19 @@ function decisionMaker(a::T1, state::T2)::Dict{Symbol, Any} where {T1<:agent, T2
   }
 
   Here are some examples:
-  {
-    "Question": "I would like to buy a sedan.",
-    "Thought_1": "I have many cars in my inventory suitable for several usage scenarios.",
-    "Thought_2": "It would be better if I knew what the user intends to do with his car.",
-    "Thought_3": "I will ask the user what is the intended usecase",
-    "Action_1": {"name": "chatbox", "input": "What will you use it for?"}
-  }
   {
     "Question": "I'm looking for a sedan with an automatic driving feature.",
     "Thought_1": "I have many types of sedans in my inventory, each with diverse features.",
     "Thought_2": "But there is only 1 car that has the feature customer wanted.",
     "Action_1": {"name": "finish", "input": "I recommend a Tesla model Y. It has your requested feature and much more."}
   }
+  {
+    "Question": "I would like to buy a sedan with 8 seats.",
+    "Thought_1": "I have one model that fits the user demand",
+    "Thought_2": "But I'm not sure that we have it in stock.",
+    "Thought_3": "I need to check out inventory first.",
+    "Action_1": {"name": "inventory", "input": "Yiem model A"}
+  }
 
   $reflect
 

src/mcts.jl

@@ -48,7 +48,7 @@ struct MCTSNode{T<:AbstractDict}
   state::T
   visits::Integer
   stateValue::AbstractFloat
-  children::Dict{T, MCTSNode}
+  children::Dict{String, MCTSNode}
 end
 
 """ Select a node based on UCT score
@@ -121,8 +121,7 @@ function expand(a::T1, node::MCTSNode, state::T2, decisionMaker::Function,
     newStatekey, newstate = MCTStransition(a, node.state, thoughtDict)  #[] Implement your transition function
     
     if newStatekey ∉ keys(node.children)# BUG should be "key of the newstate" here not newstate itself
-      statetype = typeof(state) 
+      node.children[newStatekey] = MCTSNode(newstate, 0, 0.0, Dict{String, MCTSNode}())
-      node.children[newStatekey] = MCTSNode(newstate, 0, 0.0, Dict{statetype, MCTSNode}())
     end
 
     # add stateValueEstimator
@@ -262,7 +261,7 @@ function MCTStransition(a::T1, state::T2,
   latestObservationKey = Symbol("Observation_$(latestActionIndice)")
   newstate[:thoughtHistory][latestObservationKey] = response
 
-  newStatekey = Symbol(GeneralUtils.uuid4snakecase())
+  newStatekey = GeneralUtils.uuid4snakecase()
 
   return newStatekey, newstate
 end
@@ -370,8 +369,7 @@ function runMCTS(
   maxIterations::Integer, 
   w::Float64) where {T1<:agent}
   
-  statetype = typeof(initialState)
+  root = MCTSNode(initialState, 0, 0.0, Dict{String, MCTSNode}())
-  root = MCTSNode(initialState, 0, 0.0, Dict{statetype, MCTSNode}())
   
   for _ in 1:maxIterations
     node = root
@@ -381,7 +379,8 @@ function runMCTS(
 
     expand(a, node, node.state, decisionMaker, stateValueEstimator, n=n)
     
-    leaf_node = node.children[node.state] # mark leaf node
+    # from paper, just start simulation at this node. Not the node that newly expanded
+    leaf_node = node  
     reward = simulate(leaf_node.state, maxDepth)
     backpropagate(leaf_node, reward)
   end