GOAM: Game-Oriented Agentic Modeling for Turn-Based Game AI
License: CC BY 4.0 License Keywords: Artificial Intelligence and Machine Learning, turn-based game AI, agentic architecture, dual-process reasoning, certainty-based routing, large language models
Abstract
We present GOAM (Game-Oriented Agentic Modeling), an architecture for turn-based game AI that separates strategic orchestration from tactical execution under an explicit certainty-based routing scheme. A macro process maintains longer-horizon objectives and game context; a micro process handles immediate action execution, cached patterns, and reactive play. Routing between these layers is governed by a certainty score that limits expensive language-model deliberation to genuinely ambiguous positions, reducing cost while preserving strategic coherence across a full game trajectory.
We formalize the GOAM architecture; its macro/micro decomposition, certainty model, decision routing, and feedback structure, and examine a derivative implementation deployed as an LLM bridge for turn-based strategy. Supporting supplementary materials report isolated micro-agent results via NOESIS-based experiments in chess and tic-tac-toe. Together, these contributions position GOAM as a principled agentic architecture for turn-based game AI and motivate a broader empirical validation program.
References & Links
Supplementary Note
The GOAM supplement includes two NOESIS verification artifacts: a tic-tac-toe package and a chess package. The tic-tac-toe side is the cleaner proof of concept: with center/corner-biased self-play and a compact pattern cache, it only needed roughly 700 learned patterns to force optimal play. Small self-play slices can still show temporary X/O imbalance because of first-move initiative and nonzero temperature, but the intended interpretation is that Noesis should be effectively even against itself once the pattern regime stabilizes. The chess side should be read differently. It demonstrates NOESIS as a fast pattern-oracle for instinctive move selection and certainty scoring, not as an exhaustive “solved chess” claim; the more meaningful signal is a draw-heavy self-play regime with occasional white/black wins depending on stopping point and temperature settings. Earlier pre-validation chess runs, where move legality was not yet enforced correctly, should be treated as debugging artifacts rather than reported results.