algorhino-anemone 0.1.19

anemone searches trees

anemone

anemone is a Python library for tree search over valanga game states. It builds a shared tree graph and layers algorithm-specific wrappers on top so you can plug in node evaluation, exploration indices, and selection policies for "tree and value" searches.

Highlights

• Tree-and-value exploration pipeline driven by TreeAndValueBranchSelector.
• Modular factories for node evaluation, selection, index computation, and tree management.
• Pluggable stopping criteria and recommender rules for final branch selection.
• Optional torch-based evaluator for batched neural evaluations.

Installation

pip install anemone

Optional torch integration:

pip install anemone[nn]

Quick start

anemone exposes factory helpers to build a branch selector configured with your node selector, evaluation, and stopping-criterion choices. At runtime you feed it a valanga state and a seed to get back a branch recommendation.

from random import Random

from anemone import TreeAndValuePlayerArgs, create_tree_and_value_branch_selector
from anemone.node_selector.factory import UniformArgs
from anemone.node_selector.node_selector_types import NodeSelectorType
from anemone.progress_monitor.progress_monitor import (
    StoppingCriterionTypes,
    TreeBranchLimitArgs,
)
from anemone.recommender_rule.recommender_rule import SoftmaxRule

# Populate the pieces specific to your game domain.
args = TreeAndValuePlayerArgs(
    node_selector=UniformArgs(type=NodeSelectorType.UNIFORM),
    opening_type=None,
    stopping_criterion=TreeBranchLimitArgs(
        type=StoppingCriterionTypes.TREE_BRANCH_LIMIT,
        tree_branch_limit=100,
    ),
    recommender_rule=SoftmaxRule(type="softmax", temperature=1.0),
)

selector = create_tree_and_value_branch_selector(
    state_type=YourStateType,
    args=args,
    random_generator=Random(0),
    master_state_evaluator=your_state_evaluator,
    state_representation_factory=None,
    queue_progress_player=None,
)

recommendation = selector.select_branch(state=current_state, selection_seed=0)
print(recommendation.branch_key)

Design

This codebase follows a “core node + wrappers” pattern.

• TreeNode (core)

  • TreeNode is the canonical, shared data structure.
  • It stores the graph structure: branches_children and parent_nodes.
  • There is conceptually a single tree/graph of TreeNodes.

• Wrappers implement ITreeNode

  • Higher-level nodes (e.g. AlgorithmNode) wrap a TreeNode and add algorithm-specific state: evaluation, indices, representations, etc.
  • Wrappers expose navigation by delegating to the underlying TreeNode.

• Homogeneity at the wrapper level

  • Even though TreeNode is the core place where connections are stored, each wrapper is intended to be closed under parent/child links:
    • a wrapper’s branches_children and parent_nodes contain that same wrapper type.
    • today this is typically either “all TreeNode” or “all AlgorithmNode”.
    • in the future, another wrapper can exist (still implementing ITreeNode), and it should also be homogeneous within itself.

The practical motivation is:

• algorithms can be written against ITreeNode (for navigation) and against wrappers like AlgorithmNode (for algorithm-specific fields),
• while keeping a single shared underlying structure that can be accessed consistently from any wrapper.

Repository layout

Each important package folder includes a local README with details. Start with:

• src/anemone/ for the main search pipeline and public entry points.
• src/anemone/node_selector/ for selection strategies (Uniform, RecurZipf, Sequool).
• src/anemone/node_evaluation/ for direct evaluation and minmax tree evaluation.
• src/anemone/tree_manager/, src/anemone/trees/, and src/anemone/updates/ for tree construction, expansion, and backpropagation.
• src/anemone/indices/ for exploration index computation and updates.
• tests/ for index and tree-building fixtures.