Building a 9-ball AI player: Candidate generation for direct cut shots [P]

2026-05-04 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, medium

Summary

A developer is building an AI player for 9-ball pool to optimize pattern play by selecting shots based on success probability and favorable subsequent positions. The system integrates a Transformer-based model to predict win probability from table layouts, a candidate shot generator for various shot types (cut, bank, kick, carom, combination, safeties), and an evaluator that uses the win probability model. To overcome the computational expense of the `pooltool` physics simulator, which takes 5-15 ms per shot, the developer implemented a faster evaluation method. This involves pre-computed "Acceptance window" and "Shot-index" lookup tables, complemented by a small, 50k-parameter MLP "throw model" that generalizes shot data into continuous space. This approach achieves a 10,000x speedup for candidate evaluation, reducing 1000 candidate shots from 10 seconds to 1 ms, and an overall 50-100x speedup for end-to-end shot selection.

Key takeaway

For AI Engineers developing agents in physics-rich environments, consider a hybrid approach that combines pre-computed physics lookups and small, specialized ML models with full physics simulations. Your team can achieve significant computational speedups (e.g., 50-100x end-to-end) by using ML for rapid candidate generation and filtering, reserving expensive physics simulations only for final validation of promising options. This strategy allows for extensive self-play data generation and faster model iteration.

Key insights

Hybrid AI systems combining physics simulation with ML models can drastically accelerate complex decision-making.

Principles

• Decompose complex actions into sub-problems.
• Pre-compute static physics for dynamic systems.
• Use ML to generalize discrete physics data.

Method

The method involves pre-computing shot acceptance windows and shot indices, then training a small MLP "throw model" to generalize shot parameters. This enables rapid batch evaluation of candidate shots on GPUs, followed by physics-based validation of top candidates.

In practice

• Pre-compute lookup tables for fixed physics.
• Train small MLPs for continuous generalization.
• Use noisy simulations for robust shot selection.

Topics

• 9-ball AI Player
• Candidate Shot Generation
• Billiards Simulation
• Transformer Model
• Multilayer Perceptron

Best for: Machine Learning Engineer, AI Engineer, AI Scientist

Related on AIssential

• A Model-Driven Approach for Developing Families of Reinforcement Learning Environments — A model-driven approach using a hybrid genetic algorithm automates the creation of diverse reinforcement learning environment families.
• Building a Chess Coach — Anant Dole and Asbjorn Steinskog, Take Take Take — Combining classical chess engines with LLMs and human-like move prediction enhances AI chess coaching.
• [P] Vibecoded on a home PC: building a ~2700 Elo browser-playable neural chess engine with a Karpathy-inspired AI-assisted research loop — An AI-assisted research loop can rapidly develop high-performance neural chess engines on consumer hardware.
• This Forgotten Idea Is Taking Over Computing — Stochastic computing harnesses inherent physical randomness to perform calculations, offering an energy-efficient alternative for specific AI workloads.
• Efficiently Linking Real Scenes with Synthetic Data Generation for AI-based Cognitive Robotics and Computer Vision Applications — Bridging the sim-to-real domain gap via linked synthetic and real data is crucial for scalable AI robotics.
• What Parameter Golf taught us about AI-assisted research — AI coding agents significantly impact ML competitions by lowering barriers and accelerating experimentation, while posing new review challenges.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.