A Model-Driven Generative Self Play-Based Toolchain for Developing Games and Players
summary
A model-driven toolchain generates an executable interpreter from a game description language (Stanford GDL-based with extensions for random events, hidden information, simultaneous moves, arithmetic and a type system), and a neural-net DSL Montana that emits MXNet/TensorFlow models. The interpreter becomes an RL environment for iterative self-play (copy-and-improve loop). Evaluated on tic‑tac‑toe, chess and a card game; DQN often learns legal play but stalls at draws, while other RL variants and population training give better players.
AI-written summary · about
chapters (8)
- 0:00Problem domain intro
- 1:06Toolchain overview
- 2:30Game Description Language
- 6:08Neural network integration
- 7:56Reinforcement learning training
- 8:59Evaluation with games
- 10:47Achievements and future work
- 12:15Q&A
related talks
39mLet LLMs Wander: Engineering RL Environments
46mAI Takeover
42mReinforcement Learning using Generative Models for Continuous State and Action Space Systems
15mAI Self - Learning Game Playing
01h 03mIEEE CoG 2019 - Industry Day - David Silver (DeepMind)
31mDesigning game AI with reinforcement learning
47mHow we use Reinforcement Learning (...)
01h 17mTraining intelligent game agents using deep reinforcement learning