Predicting Competitive Pokémon VGC Leads Using Latent Semantic Analysis
10 months ago
- #Latent Semantic Analysis
- #Machine Learning
- #Pokémon VGC
- The study explores using Latent Semantic Analysis (LSA) to predict Pokémon leads in competitive VGC battles.
- Data was collected from over 5,000 Pokémon Showdown battle logs, focusing on team compositions and lead choices.
- The model was evaluated using the NAIC 2025 Top 8 bracket, achieving 62.50% accuracy in hard predictions (both leads correct) and 81.25% in soft predictions (at least one lead correct).
- Lead selection is crucial in VGC as it sets the tone for the match, with synergy and threat coverage being key factors.
- The methodology involved filtering logs to include only teams from the NAIC 2025 Top 8, resulting in a refined dataset of 1,174 battles.
- The model uses cosine similarity to predict leads based on team compositions, with performance improving as the number of predicted leads increases.
- Future work includes incorporating moveset and item data, predicting all four Pokémon selections, and optimizing team compositions for better coverage and synergy.