How FTM GAMES incorporates artificial intelligence for matchmaking
At its core, FTM GAMES leverages a sophisticated, multi-layered artificial intelligence system for matchmaking that goes far beyond simple connection speed or player level. The primary objective is to create fair, competitive, and engaging matches by analyzing a complex web of player data in real-time. This isn’t just about finding ten players; it’s about predicting the quality of the interaction they will have, aiming to maximize player retention and satisfaction by minimizing frustrating mismatches.
The system’s foundation is built on a machine learning model trained on petabytes of historical match data. This model doesn’t just look at a player’s win-loss record. It ingests a vast array of features, each weighted dynamically based on the specific game mode and current player population behavior. For instance, the importance of a “teamwork” metric might be heavily weighted in an objective-based game mode compared to a free-for-all deathmatch. The AI continuously refines these weights through A/B testing, where slight variations in the matchmaking logic are deployed to small segments of the population to measure their impact on key performance indicators (KPIs) like session length and re-queue rates.
Let’s break down the specific data points the AI analyzes before forming a lobby:
Core Player Metrics (The Obvious Stuff):
- Skill Rating (MMR): A dynamic, confidence-based score that changes more rapidly after a new player’s matches until their skill level is accurately pinpointed.
- Recent Performance: A player on a winning or losing streak may have their MMR adjusted with a volatility factor to faster place them in appropriately challenging matches.
- Ping/Latency: The system prioritizes sub-35ms ping for core gameplay servers, but may relax this constraint during off-peak hours to avoid excessively long queue times, using a trade-off algorithm.
Advanced Behavioral Metrics (The AI Secret Sauce):
- Playstyle Archetype: Players are classified into archetypes like “Aggressive Rusher,” “Strategic Sniper,” or “Support Specialist” based on in-game actions per minute, positioning heatmaps, and equipment usage. The AI strives for balanced teams with complementary archetypes.
- Communication & Social Data: Does a player use voice chat? Do they frequently use the “Thank You” quick-chat? Players with positive social scores are subtly prioritized to be matched together, reducing toxic interactions.
- Time-of-Day & Day-of-Week Patterns: The system recognizes that player skill and patience can vary depending on when they play. The matchmaking tolerance (the range of skill levels it will consider for a match) is wider during late-night hours when fewer players are online.
- Party vs. Solo Dynamics: This is a critical area. The AI applies a “party bonus” to the combined MMR of a pre-made team to account for their inherent coordination advantage over a group of solo players. The size of this bonus is itself determined by AI, analyzing the win rate of parties of different sizes and skill disparities.
The following table illustrates how the AI might evaluate two potential team compositions for a 5v5 match, showcasing the trade-offs it makes. Each factor is scored on a normalized scale from 1-10.
| Matchmaking Factor | Potential Team Comp A | Potential Team Comp B | AI’s Notes |
|---|---|---|---|
| Average Skill MMR Balance | 9.2 (Near Perfect) | 8.1 (Good) | Comp A is technically more balanced. |
| Team Archetype Diversity | 4.0 (Poor – 4 Rushers) | 9.5 (Excellent – Balanced) | Comp B has a much healthier mix of roles. |
| Predicted Ping (Avg.) | 8.5 (Good) | 9.8 (Excellent) | Comp B has a significant latency advantage. |
| Solo/Party Fairness | 3.0 (Poor – 5-stack vs. Solos) | 10.0 (Perfect – All Solos) | Comp A creates a potentially unfair social dynamic. |
| Overall Match Quality Score | 6.2 | 9.1 | The AI selects Team Comp B, favoring better long-term player experience over raw MMR parity. |
Once a match is underway, the AI’s job isn’t over. It enters an analysis phase, tracking the actual outcome against its predictions. If a match was a landslide victory, the AI dissects why. Was it due to an unaccounted-for skill discrepancy? Or was it because one team’s archetypes were hard-countered by the other’s? This post-game analysis is fed back into the model as reinforcement learning, creating a virtuous cycle of improvement. Over thousands of matches, the system learns, for example, that on a specific map, “Strategic Snipers” have a 15% higher win rate, and will adjust its pre-game team balancing to compensate for this map-specific meta.
Another layer involves handling queue times. The AI operates on a dynamic “search expansion” protocol. For the first 15 seconds in a queue, it looks for a near-perfect match within a tight MMR range. After that window, it begins to slowly expand the acceptable skill range, the ping tolerance, and the archetype flexibility. The rate of this expansion is not linear; it’s based on a real-time calculation of how many players in compatible skill brackets are currently searching in nearby server regions. The goal is always to find the best possible match before player patience runs out, a metric the studio constantly monitors. Data shows that after 90 seconds, the likelihood of a player canceling their queue increases by 300%, so the AI is engineered to make a high-quality match well before that threshold for the vast majority of players.
Furthermore, the system incorporates a form of “graceful degradation” for high-skill players. In the top 0.1% of the player base, where the number of concurrent players is small, the AI cannot always create a perfectly balanced match without imposing hour-long queues. Here, it employs a different strategy. It may create matches with a larger skill spread but introduces a “underdog bonus” system, where the lower-rated team gains significantly more MMR for a win and loses less for a defeat. This mechanic, calculated on the fly by the AI, ensures that even seemingly unbalanced matches at the highest level remain meaningful and rewarding for all participants, protecting the competitive integrity of the ranked ladder.
The infrastructure supporting this is a marvel in itself. The matchmaking AI runs on a distributed cloud computing network, allowing it to evaluate millions of potential player combinations across global servers simultaneously. The decision to form a lobby is not made by a single server but is a consensus reached by a cluster of nodes analyzing the same pool of players from slightly different angles to ensure robustness. This also provides redundancy; if one node fails, the matchmaking process continues seamlessly without players noticing a disruption. The entire system processes over 5 terabytes of new match data daily, which is used to retrain the core models weekly, ensuring the AI’s understanding of the game’s meta and player behavior is never more than a few days old.