This commit is contained in:
narawat lamaiin
2024-06-01 08:20:39 +07:00
parent 1c863cd8ca
commit abefbaa743

View File

@@ -1,729 +0,0 @@
""" https://www.harrycodes.com/blog/monte-carlo-tree-search
"""
module mcts
export MCTSNode, runMCTS, isleaf, selectBestNextState, selectBestTrajectory, transition,
userChatbox, makeNewState
using Dates, UUIDs, DataStructures, JSON3, Random, PrettyPrinting
using GeneralUtils
using ..type, ..llmfunction
# ---------------------------------------------- 100 --------------------------------------------- #
""" a node for MCTS search tree
# Arguments
- `state::T`
a state of a game. Can be a Dict or something else.
- `visits::Integer `
number of time the game visits this state
- `stateValue::Float64`
state value
- `children::Dict{T, MCTSNode}`
children node
# Return
- `nothing`
# Example
```jldoctest
julia> state = Dict(
:info=> Dict(), # keyword info
:thoughtHistory=> Dict(
:question=> _,
:thought_1=> _,
:action_1=> _,
:observation_1=> _,
:thought_2=> _,
...
)
)
```
# TODO
[] update docstring
# Signature
"""
mutable struct MCTSNode{T1<:AbstractDict, T2<:AbstractString}
nodekey::T2
state::T1
visits::Integer
progressvalue::Number # estimate value by LLM's reasoning
statevalue::Number # store discounted commulative reward (gather from its child node)
reward::Number # this node's own reward
isterminal::Bool
parent::Union{MCTSNode, Nothing}
children::Dict{String, MCTSNode}
end
""" Select a node based on UCT score
# Arguments
- `node::MCTSNode`
mcts node
- `w::T`
exploration weight. Value is usually between 1 to 2.
Value 1.0 makes MCTS balance between exploration and exploitation like 50%-50%.
Value 2.0 makes MCTS aggressively search the tree.
# Return
- `selectedNode::MCTSNode`
# Example
```jldoctest
julia>
```
# Signature
"""
function UCTselect(node::MCTSNode, w::T)::MCTSNode where {T<:AbstractFloat}
maxUCT = -Inf
selectedNode = nothing
for (childState, childNode) in node.children
UCTvalue =
if childNode.visits != 0
weightedterm = w * sqrt(log(node.visits) / childNode.visits) # explore term
childNode.statevalue + weightedterm
else # node.visits == 0 makes sqrt() in explore term error
childNode.progressvalue # exploit term
end
if UCTvalue > maxUCT
maxUCT = UCTvalue
selectedNode = childNode
end
end
return selectedNode
end
""" Expand selected node
# Arguments
- `a::T1`
One of YiemAgent's agent
- `node::MCTSNode`
MCTS node
- `state::T2`
a state of a game. Can be a Dict or something else.
- `decisionMaker::Function`
a function that output Thought and Action
- `evaluator::Function`
a function that output trajectory progress score
# Return
# Example
```jldoctest
julia>
```
# TODO
[] update docstring
[] try loop should limit to 3 times. if not succeed, skip
[] 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.
[x] store feedback -> state -> agent.
# Signature
"""
function expand(a::T1, node::MCTSNode, decisionMaker::Function,
evaluator::Function, reflector::Function; totalsample::Integer=3
) where {T1<:agent}
nthSample = 0
while true
nthSample += 1
if nthSample <= totalsample
thoughtDict = decisionMaker(a, node.state)
println("---> expand() sample $nthSample")
pprintln(node.state[:thoughtHistory])
pprintln(thoughtDict)
newNodeKey, newstate = MCTStransition(a, node.state, thoughtDict)
stateevaluation, progressvalue = evaluator(a, newstate)
if newstate[:reward] < 0
pprint(newstate[:thoughtHistory])
newstate[:evaluation] = stateevaluation
newstate[:lesson] = reflector(a, newstate)
# store new lesson for later use
lessonDict = copy(JSON3.read("lesson.json"))
latestLessonKey, latestLessonIndice =
GeneralUtils.findHighestIndexKey(lessonDict, "lesson")
nextIndice = latestLessonKey == :NA ? 1 : latestLessonIndice + 1
newLessonKey = Symbol("lesson_$(nextIndice)")
lessonDict[newLessonKey] = newstate
open("lesson.json", "w") do io
JSON3.pretty(io, lessonDict)
end
print("---> reflector()")
end
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
# Arguments
- `a::T`
one of YiemAgent's agent
- `node::MCTSNode`
node that will be a simulation starting point.
- `decisionMaker::Function`
function that receive state return Thought and Action
# Return
- `simTrajectoryReward::Number`
# Example
```jldoctest
julia>
```
# TODO
- [] update docs
# Signature
"""
function simulate(a::T, node::MCTSNode, decisionMaker::Function, evaluator::Function,
reflector::Function; maxDepth::Integer=3, totalsample::Integer=3
)::Union{Tuple{Number, Dict{Symbol, <:Any}}, Tuple{Number, Nothing}} where {T<:agent}
simTrajectoryReward = 0.0
terminalstate = nothing
for depth in 1:maxDepth
simTrajectoryReward += node.reward
if node.isterminal
terminalstate = node.state
break
else
expand(a, node, decisionMaker, evaluator, reflector; totalsample=totalsample)
node = selectChildNode(node)
end
end
return (simTrajectoryReward, terminalstate)
end
""" Backpropagate reward along the simulation chain
# Arguments
- `node::MCTSNode`
leaf node of a search tree
- `simTrajectoryReward::T`
total reward from trajectory simulation
# Return
- `No return`
# Example
```jldoctest
julia>
```
# Signature
"""
function backpropagate(node::MCTSNode, simTrajectoryReward::T;
discountRewardCoeff::AbstractFloat=0.9) where {T<:Number}
while !isroot(node)
# Update the statistics of the current node based on the result of the playout
node.visits += 1
node.statevalue += ((node.statevalue * (node.visits-1)) + simTrajectoryReward) / node.visits
simTrajectoryReward *= discountRewardCoeff # discount because future reward is uncertain
node = node.parent
end
end
""" Get a new state
# Arguments
- `a::T1`
one of YiemAgent's agent
- `state::T2`
current game state
- `thoughtDict::T3`
contain Thought, Action, Observation
- `isterminal::Function`
a function to determine terminal state
# Return
- `(newNodeKey, newstate, isterminalstate, reward)::Tuple{String, Dict{Symbol, <:Any}, Bool, <:Number}`
# Example
```jldoctest
julia> state = Dict{Symbol, Dict{Symbol, Any}}(
:thoughtHistory => Dict(:question => "Hello, I want to buy a bottle of wine."),
:storeinfo => Dict(),
:customerinfo => Dict()
)
julia> thoughtDict = Dict(
:question=> "I want to buy a bottle of wine.",
:thought_1=> "The customer wants to buy a bottle of wine.",
:action_1=> Dict{Symbol, Any}(
:name=>"Chatbox",
:input=>"What occasion are you buying the wine for?",
),
:observation_1 => ""
)
```
# TODO
- [x] add other actions
- [] add embedding of newstate and store in newstate[:embedding]
# Signature
"""
function MCTStransition(a::T1, state::T2, thoughtDict::T2
)::Tuple{String, Dict{Symbol, <:Any}} where {T1<:agent, T2<:AbstractDict}
actionname = thoughtDict[:action][:name]
actioninput = thoughtDict[:action][:input]
# map action and input() to llm function
response, select, reward, isterminal =
if actionname == "chatbox"
# deepcopy(state[:virtualCustomerChatHistory]) because I want to keep it clean
# so that other simulation start from this same node is not contaminated with actioninput
virtualWineUserChatbox(a, actioninput, deepcopy(state[:virtualCustomerChatHistory])) # virtual customer
elseif actionname == "winestock"
winestock(a, actioninput)
elseif actionname == "recommendbox"
virtualWineUserRecommendbox(a, actioninput)
else
error("undefined LLM function. Requesting $actionname")
end
newNodeKey, newstate = makeNewState(state, thoughtDict, response, select, reward, isterminal)
if actionname == "chatbox"
push!(newstate[:virtualCustomerChatHistory], Dict(:name=>"assistant", :text=> actioninput) )
push!(newstate[:virtualCustomerChatHistory], Dict(:name=>"user", :text=> response))
end
return (newNodeKey, newstate)
end
""" Get a new state
# Arguments
- `a::T1`
one of YiemAgent's agent
- `state::T2`
current game state
- `thoughtDict::T3`
contain Thought, Action, Observation
- `isterminal::Function`
a function to determine terminal state
# Return
- `(newNodeKey, newstate, isterminalstate, reward)::Tuple{String, Dict{Symbol, <:Any}, Bool, <:Number}`
# Example
```jldoctest
julia> state = Dict{Symbol, Dict{Symbol, Any}}(
:thoughtHistory => Dict(:question => "Hello, I want to buy a bottle of wine."),
:storeinfo => Dict(),
:customerinfo => Dict()
)
julia> thoughtDict = Dict(
:question=> "I want to buy a bottle of wine.",
:thought_1=> "The customer wants to buy a bottle of wine.",
:action_1=> Dict{Symbol, Any}(
:name=>"Chatbox",
:input=>"What occasion are you buying the wine for?",
),
:observation_1 => ""
)
```
# TODO
- [x] add other actions
- [] add embedding of newstate and store in newstate[:embedding]
# Signature
"""
function transition(a::T1, state::T2, thoughtDict::T2
)::Dict{Symbol, <:Any} where {T1<:agent, T2<:AbstractDict}
thoughtDict = state[:thoughtDict]
actionname = thoughtDict[:action][:name]
actioninput = thoughtDict[:action][:input]
# map action and input() to llm function
response, select, reward, isterminal =
if actionname == "winestock"
winestock(a, actioninput)
else
error("undefined LLM function. Requesting $actionname")
end
return makeNewState(state, thoughtDict, response, select, reward, isterminal)
end
"""
# Arguments
# Return
# Example
```jldoctest
julia>
```
# TODO
- [] update docstring
- [x] implement the function
# Signature
"""
function makeNewState(currentstate::T1, thoughtDict::T4, response::T2, select::Union{T3, Nothing},
reward::T3, isterminal::Bool
)::Tuple{String, Dict{Symbol, <:Any}} where {T1<:AbstractDict, T2<:AbstractString, T3<:Number, T4<:AbstractDict}
currentstate_latestThoughtKey, currentstate_latestThoughtIndice =
GeneralUtils.findHighestIndexKey(currentstate[:thoughtHistory], "thought")
currentstate_nextIndice = currentstate_latestThoughtKey == :NA ? 1 : currentstate_latestThoughtIndice + 1
currentstate_latestThoughtKey = Symbol("thought_$currentstate_nextIndice")
latestActionKey = Symbol("action_$currentstate_nextIndice")
_, thoughtDict_latestThoughtIndice =
GeneralUtils.findHighestIndexKey(thoughtDict, "thought")
thoughtDict_latestThoughtKey, thoughtDict_latestActionKey =
if thoughtDict_latestThoughtIndice == -1
(:thought, :action)
else
(
Symbol("thought_$thoughtDict_latestThoughtIndice"),
Symbol("action_$thoughtDict_latestThoughtIndice"),
)
end
# add Thought, action, observation to thoughtHistory
newstate = deepcopy(currentstate)
newstate[:thoughtHistory][currentstate_latestThoughtKey] =
thoughtDict[thoughtDict_latestThoughtKey]
newstate[:thoughtHistory][latestActionKey] = thoughtDict[thoughtDict_latestActionKey]
newObservationKey = Symbol("observation_$(currentstate_nextIndice)")
newstate[:thoughtHistory][newObservationKey] = response
newstate[:reward] = reward
newstate[:select] = select
newstate[:isterminal] = isterminal
newNodeKey = GeneralUtils.uuid4snakecase()
return (newNodeKey, newstate)
end
""" Determine whether a node is a leaf node of a search tree.
# Arguments
- `node::MCTSNode`
a search tree node
# Return
- `result::Bool`
true if it is a leaf node, false otherwise.
# Example
```jldoctest
julia> using Revise
julia> using YiemAgent, DataStructures
julia> initialState = Dict{Symbol, Any}(
:customerinfo=> Dict{Symbol, Any}(),
:storeinfo=> Dict{Symbol, Any}(),
:thoughtHistory=> OrderedDict{Symbol, Any}(
:question=> "How are you?",
)
)
julia> statetype = typeof(initialState)
julia> root = YiemAgent.MCTSNode(initialState, 0, 0.0, Dict{statetype, YiemAgent.MCTSNode}())
julia> YiemAgent.isleaf(root)
true
```
# TODO
[] update docs
# Signature
"""
isleaf(node::MCTSNode)::Bool = isempty(node.children)
""" Select child node based on the highest statevalue
# Arguments
- `node::MCTSNode`
node of a search tree
# Return
- `childNode::MCTSNode`
the highest value child node
# Example
```jldoctest
julia>
```
# Signature
"""
function selectChildNode(node::MCTSNode)::MCTSNode
highestProgressValue = 0
nodekey = nothing
# loop thought node children dictionary to find the highest progress value
for (k, childNode) in node.children
potential = childNode.progressvalue + childNode.reward
if childNode.reward > 0 #XXX for testing. remove when done.
println("")
end
if potential > highestProgressValue
highestProgressValue = potential
nodekey = childNode.nodekey
end
end
return node.children[nodekey]
end
"""
# Arguments
- `node::MCTSNode`
node of a search tree
# Return
- `childNode::MCTSNode`
the highest value child node
# Example
```jldoctest
julia>
```
# TODO
- [] update docs
- [x] implement the function
# Signature
"""
function selectBestNextState(node::MCTSNode)::MCTSNode
highestProgressValue = 0
nodekey = nothing
# if all childnode has statevalue == 0, use progressvalue + reward to select the best node
stateValueSum = sum([v.statevalue for (k, v) in node.children])
if stateValueSum != 0
for (k, childnode) in node.children
potential = childnode.statevalue / childnode.visits
if potential > highestProgressValue
highestProgressValue = potential
nodekey = childnode.nodekey
end
end
else
for (k, childnode) in node.children
potential = childnode.progressvalue + childnode.reward
if potential > highestProgressValue
highestProgressValue = potential
nodekey = childnode.nodekey
end
end
end
return node.children[nodekey]
end
"""
# Arguments
- `node::MCTSNode`
node of a search tree
# Return
- `childNode::MCTSNode`
the highest value child node
# Example
```jldoctest
julia>
```
# TODO
- [] update docs
- [x] implement the function
# Signature
"""
function selectBestTrajectory(node::MCTSNode)::MCTSNode
while !isleaf(node)
node = selectBestNextState(node)
end
return node
end
""" Determine wheter a given node is a root node
# Arguments
- `node::MCTSNode`
node of a search tree
# Return
- `isrootnode::Bool`
true if the given node is root node, false otherwise
# Example
```jldoctest
julia>
```
# Signature
"""
isroot(node::MCTSNode)::Bool = node.nodekey == "root" ? true : false
# ------------------------------------------------------------------------------------------------ #
# Create a complete example using the defined MCTS functions #
# ------------------------------------------------------------------------------------------------ #
""" Search the best action to take for a given state and task
# Arguments
- `a::agent`
one of Yiem's agents
- `initial state`
initial state
- `decisionMaker::Function`
decide what action to take
- `evaluator::Function`
assess the value of the state
- `reflector::Function`
generate lesson from trajectory and reward
- `isterminal::Function`
determine whether a given state is a terminal state
- `n::Integer`
how many times action will be sampled from decisionMaker
- `w::Float64`
exploration weight. Value is usually between 1 to 2.
Value 1.0 makes MCTS balance between exploration and exploitation like 50%-50%
Value 2.0 makes MCTS aggressively search the tree
# Return
- `plan::Vector{Dict}`
best plan
# Example
```jldoctest
julia>
```
# TODO
[] update docstring
[x] return best action
# Signature
"""
function runMCTS(
a::T1,
initialState,
decisionMaker::Function,
evaluator::Function,
reflector::Function;
totalsample::Integer=3,
maxDepth::Integer=3,
maxiterations::Integer=10,
explorationweight::Number=1.0,
) where {T1<:agent}
root = MCTSNode("root", initialState, 0, 0, 0, 0, false, nothing, Dict{String, MCTSNode}())
for nth in 1:maxiterations
node = root
node.visits += 1
while !isleaf(node)
node = UCTselect(node, explorationweight)
end
if node.isterminal
# MCTS arrive at the leaf node that is also a terminal state,
# do nothing then go directly to backpropagation
backpropagate(leafNode, node.reward)
else
expand(a, node, decisionMaker, evaluator, reflector; totalsample=totalsample)
leafNode = selectChildNode(node)
simTrajectoryReward, terminalstate = simulate(a, leafNode, decisionMaker, evaluator,
reflector; maxDepth=maxDepth, totalsample=totalsample)
if terminalstate !== nothing #XXX not sure why I need this
terminalstate[:totalTrajectoryReward] = simTrajectoryReward
end
#[] write best state to file if it has higher simTrajectoryReward. Use to improve evaluation
# open("trajectory.json", "w") do io
# JSON3.pretty(io, terminalstate)
# end
backpropagate(leafNode, simTrajectoryReward)
end
end
bestNextState = selectBestNextState(root)
besttrajectory = selectBestTrajectory(root)
return (bestNextState.state, besttrajectory.state)
end
end # module mcts