Amber Eye: a Slay the Spire engine approaching top-level play (8e9f.github.io)

🤖 AI Summary
Amber Eye has emerged as a notable AI engine specifically designed to play "Slay the Spire," showcasing impressive capabilities by winning 90 out of 100 games, including a remarkable streak of 19 consecutive wins. Utilizing AlphaZero-style training methods, the engine was developed from scratch on a personal computer over approximately one month, learning the intricacies of the game through self-play—totaling around 13.8 million games. The engine's innovative approach is significant as it attempts to optimize play without prior knowledge of the game's mechanics, focusing on an idealized version of "Slay the Spire" where random outcomes are completely stochastic, thus circumventing the limitations posed by the game's random seed. Despite its high win rate, the engine exhibits a playstyle that diverges from typical strategies, often making questionable tactical decisions yet still consistently securing victories. This paradox highlights an intriguing aspect of its learning process, suggesting that there are underlying strategies worth further exploration within the AI/ML community. The engine provides valuable insights, such as identifying the most impactful cards and choices throughout the game, which could help both players and developers understand optimal strategies and the implications of AI decision-making in gaming contexts. These findings not only support the development of more advanced AI systems but also contribute to the broader conversation about AI's role in mastering complex games.
Loading comments...
loading comments...