What inputs does mcts_expand require?

node_id, actions, states, priors; each action is an object with action, state, prior, and metadata fields.

How are priors used during expansion?

Priors estimate action quality and guide which actions are most likely to be expanded; actions with extremely low priors are pruned.

How do I decide which child to simulate next?

The output should include the number of children added and a recommendation on which child looks most promising for simulation; pick that one for the SIMULATION phase.

mcts-expand

Scanned

npx machina-cli add skill NewJerseyStyle/plugin-mcts/mcts-expand --openclaw

Files (1)

SKILL.md

2.2 KB

MCTS Expansion Phase

You are executing the EXPANSION phase of Monte Carlo Tree Search.

LLM as World Model

Use your knowledge to predict:

Valid actions from the current state
Resulting states from each action
Transition probabilities (if applicable)
Prior probabilities for action quality

Expansion Algorithm

Identify the node to expand (from selection phase)
Generate possible actions:
- What are the valid next steps?
- What knowledge/constraints apply?
- What has worked in similar situations?
Predict outcomes:
- For each action, what state results?
- What is the likelihood of success?
- What are potential issues?
Create child nodes for promising actions

Using MCP Tools

Call mcts_expand with:

node_id: The node to expand
actions: List of actions to add as children
states: Resulting states for each action
priors: Prior probability estimates for each action

Each action should include:

{
  "action": "description of action",
  "state": "resulting state description",
  "prior": 0.3,
  "metadata": {"reasoning": "why this action"}
}

Expansion Strategy

For the current context: $ARGUMENTS

For Research Problems:

Actions = different hypotheses or investigation paths
States = knowledge states after investigation
Priors = based on domain knowledge

For Planning Problems:

Actions = possible decisions or steps
States = situation after each decision
Priors = based on feasibility and expected outcomes

For Coding Problems:

Actions = implementation approaches or fixes
States = code states after changes
Priors = based on best practices and complexity

Pruning Considerations

Don't expand:

Obviously invalid actions
Duplicate states (detected via state hashing)
Actions with extremely low prior probability

Output

After expansion, report:

Number of children added
Brief description of each action
Prior probabilities and reasoning
Which child looks most promising for simulation

Proceed to SIMULATION phase with one of the new child nodes.

Source

git clone https://github.com/NewJerseyStyle/plugin-mcts/blob/main/skills/mcts-expand/SKILL.mdView on GitHub

Overview

This skill implements the EXPANSION phase of Monte Carlo Tree Search by using an LLM as the world model to propose valid actions, predict resulting states, and assign priors. It creates new child nodes for promising actions and prunes obviously invalid or duplicate states to keep the search focused.

How This Skill Works

From the selected node, generate valid next actions using the LLM as a world model, incorporating constraints and prior knowledge. For each action, predict the resulting state, estimate the likelihood of success, and surface potential issues. Assemble child candidates as action/state/prior/metadata records and call mcts_expand to materialize the new nodes, applying pruning rules to remove invalid or duplicate states.

When to Use It

Expanding a node in a research problem to explore hypotheses and knowledge states
Enumerating feasible decisions or steps in planning problems
Trying different implementation approaches in coding problems
Investigating alternate knowledge states after an action in uncertain domains
Pruning obviously invalid or low-prior actions before running simulations

Quick Start

Step 1: Identify the node to expand from the selection phase
Step 2: Generate possible actions and predicted states using the LLM world model; apply pruning rules to drop invalid or low-prior options
Step 3: Call mcts_expand with node_id, actions, states, priors; review the number of children added and which is most promising for simulation

Best Practices

Always identify the node to expand from the selection phase before generating actions
Generate only valid, constraint-aware actions (avoid obviously invalid moves)
Predict outcomes for each action and assign meaningful priors based on knowledge
Hash resulting states to detect duplicates and prune duplicates efficiently
Record concise metadata explaining why each action was chosen and which appears most promising

Example Use Cases

Expanding a game-tree node to test new move candidates with predicted outcomes
Building a hypothesis tree in a research assistant to compare investigation paths
Evaluating coding approaches by expanding a node with alternative implementations
Planning robot tasks by expanding decisions and forecasting resulting states
Assessing logistics routes by expanding options with priors and predicted states

Frequently Asked Questions

Add this skill to your agents