CTOR Learning Platform

CTOR Learning Platform is a high-performance research and engineering system for training game AI through deep reinforcement learning. It is built around CTOR (Circular TORus), a two-player strategy game on a 10x10 toroidal board where edges wrap around and tactical patterns behave differently from ordinary grid games.

AI Training

The project implements self-play training loops, experience replay, model checkpoints, auto-resume, baseline opponents, and multiple neural architectures. Later versions added phase-aware inputs, differential reward shaping, future vulnerability penalties, connectivity rewards, and model families such as ResNet-style networks and TViT-Dual.

Evaluation And Tournaments

A distributed tournament system evaluates trained models against historical checkpoints and algorithmic players such as Minimax, SmartMinimax, Defensive, and Random. It includes ELO calculations, match scheduling, worker queues, tournament metadata, replay support, and frontend dashboards for rankings, system telemetry, and infrastructure status.

Infrastructure

The platform combines Python game-engine code, PyTorch training, FastAPI services, Redis queues and events, MongoDB metadata, SQLite double-write migration, compressed NPZ game storage, Docker profiles, monitoring dashboards, and Next.js frontends. The system is designed to run local experiments, scale game workers, preserve training progress, and make model quality measurable rather than anecdotal.

Engineering Value

CTOR Learning is a compact laboratory for agentic AI engineering: a real rules engine, adversarial environment, model lifecycle, observability, distributed execution, storage migration, and UX surfaces for playing, testing, and comparing AI agents.