Subversion Gamer – Overview
Welcome to the forefront of game theory and strategic decision-making with Subversion Gamer.
Superhuman decision making.
Designed to tackle the complexities of modern games, Subversion Gamer goes beyond mere computation to understand and navigate the intricacies of strategy and competition.
A master of optimizing strategies in dynamically changing environments, Subversion Gamer is our machine learning framework utilizing state-of-the-art reinforcement learning algorithms and rigorous training regimens to produce an agent (AI) equipped to make intelligent, strategic decisions that forge paths toward unparalleled achievements across various domains, including financial markets, sports, business strategy, and games. Leveraging our team’s extensive knowledge on game theory and competition, we built Subversion Gamer to transform the landscape of strategy-centric AI.
Continue reading to see how Subversion Gamer conquered the game of No Limit Holdem poker with record breaking compute efficiency.
If there’s a correct strategy, we’re going to find it.
1. Employing cutting-edge techniques and capitalizing on recent advancements in accelerated computing, Subversion Gamer achieved mastery over No Limit Hold’em Poker—a game renowned for its strategic depth and complexity and a benchmark in the machine learning community. Subversion Gamer reached super-human skill level in just 8 hours using a single NVIDIA RTX 4090 GPU, at a total operational cost of just under 4 dollars compared to 10000+ GPU hours in previous poker AI’s (Facebook’s ReBeL). It also has a much stronger performance compared to just purely training on a reinforcement algorithm such as PPO.
Subversion Gamer achieves super-human, super-AI skill level in No Limit Holdem Poker with unprecedented efficiency, just 8 hours on a single consumer grade GPU.
2. The trained agent is capable of defeating other super-human AI like Slumbot (latest Annual Computer Poker Competition champion) by wide margins. Note that Subversion Gamer was trained on different rules (varying starting stacks etc.) and so results could be further improved by training strictly using Slumbot’s rules (200 big blind set stacks, 2-players).
3. The codebase for the vectorized poker environment, utilized by this iteration of Subversion Gamer for its training, is freely accessible on our GitHub page. This environment, leveraging the full potential of Google’s JAX library, has been optimized for speed, achieving performance rates thousands of times faster than conventional poker environments. This optimization was pivotal to Subversion Gamer’s groundbreaking success in mastering No Limit Hold’em.