Hive Neural Network: Using Neural Networks to Play Abstract Strategy Board Games

Project Basic Information

• Original Author/Maintainer: chrismejias
• Source Platform: GitHub
• Original Project Name: hive_neuralnet
• Project Link: https://github.com/chrismejias/hive_neuralnet
• Release Date: May 23, 2026

Core Objectives

Explore the application of reinforcement learning and traditional chess AI in abstract strategy games, focusing on the unique board game Hive.

Project Significance

Hive's features such as dynamic board and high branching factor bring new challenges to neural networks. The project's results can promote research in fields like dynamic structure deep learning.

Introduction to Hive Game

Hive is an abstract strategy board game designed by John Yianni in 2001, known as the "insect version of chess", with the following characteristics:

No Board Design

Start with one piece, and other pieces are placed around to form a dynamically expanding "hive" structure. Each game's layout is unique.

Piece Types

• Queen Bee: Core piece; the winning condition is to围困 the opponent's Queen Bee
• Beetle: Can climb onto other pieces to suppress their movement
• Spider: Must move three steps
• Ant: Extremely mobile, can reach any edge position
• Grasshopper: Jumps over pieces along a straight line

Victory Condition

A player loses if their Queen Bee is completely surrounded by six pieces (regardless of friend or foe).

Complexity

• State Space: Theoretically infinite (actually finite)
• Branching Factor: Average of 30-40 legal moves per step (higher than chess)
• No Randomness: Pure strategy game

Hive is simpler than Go (smaller state space) but more complex than chess (higher branching factor, dynamic board).

Challenges of Neural Networks in Chess Games

Comparison with Traditional Chess AI

Feature	Chess	Go	Hive
Board	Fixed 8×8	Fixed 19×19	Dynamically expanding
Pieces	Fixed number	No pieces	Placed gradually
State Representation	Simple matrix	Simple matrix	Complex graph structure
Game Length	~80 moves	~200 moves	~40-60 moves

Architecture Challenges

1. Input Representation: Traditional CNNs are not suitable for dynamic boards; variable-length input architectures like Graph Neural Networks (GNNs) are needed
2. Position Encoding: Only relative positions matter, requiring the network to have translation invariance
3. Topology: Hexagonal grids are not suitable for rectangular convolution kernels

Possible Implementation Schemes

Reinforcement Learning Framework

Mainstream process:

1. Self-play: Generate a large amount of game data
2. Policy Network: Select the best action
3. Value Network: Evaluate winning rate
4. Monte Carlo Tree Search (MCTS): Combine policy and value networks for search

Network Architecture Choices

• Graph Convolutional Network (GCN): Capture local topological features
• Attention Mechanism: Learn important relationships between pieces
• Transformer: Treat the board as a token sequence and use self-attention

Training Techniques

• Curriculum Learning: Gradually increase difficulty from simple positions
• Residual Learning: Predict the difference from the current strategy
• Multi-task Learning: Predict strategy and value simultaneously

Significance of Chess AI Research

1. Algorithm Test Platform: Controllable environment, clear evaluation, low simulation cost
2. Strategy Learning Paradigm: Reinforcement learning experience can be transferred to fields like game AI, robot control, and resource scheduling
3. Human Cognition Research: Compare human and AI styles to explore differences between intuition and computation

Potential Value of Hive AI

1. Game Assistance: Online battle AI, training partner, position analysis
2. Algorithm Research: Testbed for dynamic structure deep learning, hexagonal grid representation, and few-shot reinforcement learning
3. Educational Value: Case for teaching game theory and AI concepts

Related Research and Projects

AlphaZero

DeepMind's milestone project, which learns from scratch through self-play. Its core ideas (policy + value network + MCTS) have become the standard paradigm

Other Chess AIs

• Leela Chess Zero: Open-source chess engine using a similar approach to AlphaZero
• KataGo: Open-source Go engine with improvements based on AlphaZero
• OpenAI Five: Dota2 AI breakthrough, demonstrating the potential of complex multi-agent systems

Abstract Strategy Game Research

Academia has little research on abstract strategy games like Hex and Y, and Hive is even more niche. This project may fill the gap

Summary and Outlook

chrismejias's hive_neuralnet project is concise but contains rich technical challenges and research value. Hive provides a unique experimental platform for neural networks and reinforcement learning, and the experience of handling dynamic graph structures and hexagonal grids can be applied to broader AI fields.

Recommendations

• Follow the project's subsequent development
• Try starting with heuristic search and gradually introduce neural networks and reinforcement learning; it is a highly educational learning journey

Future Expectations

With the development of deep learning, it is expected to see superhuman-level Hive AI, which will become one of the milestones of general artificial intelligence