Judgion: A Frist Approach to Applying Machine Learning Techniques to Judge Mixed Martial Arts Bouts
Abstract
Ivan Ontiveros
In this work, I present Judgion, a family of six lightweight neural models for round-level MMA scoring from round-by- round statistics. The models are trained on a 501-round corpus that includes real UFC rounds and a set of theoretical rounds crafted to encode key judging criteria that real rounds may not be enough to capture. Judgion’s goal is not to predict how judges would score a fight, but to give their own scores following the judging criteria, learning the rules and applying them to real rounds. Judgion was evaluated live on UFC 317–319. Most model scorecards were defensible under the criteria; 7/52 rounds (≈ 13.5%) were flagged as misjudged on qualitative review. At the model–round level, 23/312 decisions (≈ 7.4%) were flagged. Given the complexity and subjectivity of judging and the coarseness of publicly available data, these results are promising and highlight a clear path forward: with richer inputs and continued refinement, automated round scoring can become substantially more reliable.