Researchers from the Singapore University of Technology and Design (SUTD) have achieved a breakthrough by applying reinforcement learning to solve complex video game challenges. Their innovative approach, inspired by strategies used in board games like Chess and Go, outperformed top human players in a single test.
This milestone has far-reaching implications, potentially revolutionizing movement design in fields like robotics and automation. The team’s research, published in Advanced Intelligence Systems under the title “A Phase-Change Memristive Reinforcement Learning for Rapidly Outperforming Champion Street Fighter Players,” highlights the versatility of reinforcement learning beyond simple board games.
Principal investigator Desmond Loke, Associate Professor at SUTD, stated, “Our findings prove that reinforcement learning can tackle intricate challenges in movement science. If applied to the right research problems, this method could accelerate progress in various scientific domains.”
This study represents a pivotal moment in leveraging artificial intelligence for advancing movement science studies, offering applications such as autonomous vehicles, collaborative robots, and aerial drones.
Reinforcement learning, a form of machine learning, enables computers to make decisions by experimenting, receiving feedback, and improving decision-making over time. For instance, algorithms can master chess by exploring countless moves and learning from the outcomes.
In this research, the team trained a reinforcement learning program for movement design using millions of initial motions. The program iteratively improved each move toward a specific goal, ultimately surpassing built-in AI opponents.
Associate Prof Loke emphasized the uniqueness of their approach, stating, “Our method stands out because it employs reinforcement learning to create movements that outperform top human players, something previously unattainable with other methods. It has the potential to reshape the range of movements we can generate.”
The researchers also demonstrated energy efficiency in their phase-change memory-based system, achieving hardware energy consumption remarkably lower than existing GPU systems.
The team focused on creating motions capable of swiftly defeating top human players, employing decay-based algorithms in their approach. Testing showed that AI-designed motions excelled in various aspects, including game etiquette, handling inaccurate information, reaching specific game states, and quickly defeating opponents.
In essence, this program demonstrated exceptional physical and mental attributes, leading to effective movement design. Decay-based training techniques enabled faster progress compared to conventional methods.
The researchers envision a future where their strategy enables the creation of movements, skills, and actions previously deemed impossible. Associate Prof Loke envisions expanded applications, including enhancing competitive tasks in games like Poker, Starcraft, and Jeopardy, as well as facilitating high-level training and discovery of new tactics in professional gaming.
Contributions to the study were made by SUTD researchers Shao-Xiang Go and Yu Jiang.