LLMMCTS/examples

LLMMCTS Examples

This directory contains example scripts demonstrating how to use LLMMCTS for various problem types.

Examples

1. simple_example.jl - Basic MCTS usage with a simple state transition function
2. pathfinding.jl - Grid-based pathfinding problem
3. math_problem.jl - Solving math problems using MCTS-guided reasoning
4. tool_use.jl - Coordinating with external tools (APIs, databases)
5. chess_game.jl - Game playing scenario (simplified chess-like)
6. code_generation.jl - Guiding LLM code generation
7. reasoning.jl - Multi-step reasoning with chain-of-thought
8. configuration_examples.jl - Demonstrating different MCTS configuration options

Running Examples

julia examples/simple_example.jl
julia examples/pathfinding.jl
julia examples/configuration_examples.jl

Key Concepts

State

The state is represented as a Dict{String, Any} that contains all information needed for the problem.

Transition Function

The transition function takes the current state and returns:

Dict(
    :newNodeKey => unique_id,
    :newstate => new_state_dict,
    :progressvalue => llm_estimate
)

Progress Value

progressvalue is provided by LLM reasoning and guides the search without waiting for terminal rewards.

State Value

statevalue is computed through Monte Carlo simulations and provides accurate long-term estimates.

Configuration Parameters

• maxiterations - Number of MCTS iterations (default: 10)
• explorationweight - UCT exploration weight (default: 1.0)
• maxSimulationDepth - Maximum simulation rollout depth (default: 3)
• horizontalSampleExpansionPhase - Children per expansion (default: 3)
• multithread - Enable parallel simulation (default: false)
• saveSimulatedNode - Keep simulation nodes (default: false)

See Also

• README.md - Complete package documentation
• workprocess.md - Detailed technical documentation