Engineering the First Self-Taught Tales of Tribute Agent via Deep RL and Abstract Action Selection

Overview

As artificial intelligence takes over more of day -to-day decision making, it is imperative that automated systems are capable of making intelligent, well-reasoned decisions. For decades, this quality of AI has been measured through computers’ mastery of in creasingly sophisticated strategy games. Today, a major complexity benchmark is Tales of Tribute (ToT), an enormously complicated online card game. To cope with the game’s intensity, most strong ToT agents rely on the use of pre-existing human knowledge via the insertion of heuristic rules, dramatically reducing those agents’ applicability to other domains. In contrast, this paper introduces RL -GG (“Reinforcement Learning, Good Game,” ) the world’s first self -taught agent for ToT. Through reinforcement learning, RL-GG teaches itself the game from the ground up, without the need for human knowledge, allowing it to be built with multi-domain versatility in mind. To do this, it utilizes a novel action representation technique, reminiscent of LLM token architecture, that allows it to evaluate an arbitrarily large number of actions without increasing computation time. In bot-versus- bot matches, RL-GG achieves winrates ranging from 41.5 ± 3.1% to 84.4 ± 2.3%, demonstrating the efficacy of its variety of improvements over vanilla RL and search -based techniques. Most importantly, these advancements produce a broadly applicable RL algorithm capable of learning delayed gratification (among other skills) from zero human input in a harshly complicated domain with extremely limited computing power. As a result, RL -GG simultaneously summits a notable challenge in computer science while also developing AI’s “toolbox” of abstract reasoning methods for use on real-world problems.

Competition history

• JSHS 2025 Category not listed

Resources

• National JSHS abstract book ↗

Related projects

Closest projects by meaning, across every fair and year in the corpus.