The Solo State Value Model (SSVM) is a data-driven model of a solo FNCS game. At any moment, it looks at everything about a player's situation and estimates how many points that player will finish the game with. We call that number their expected points. You can watch the predictions for an example game beside a map of the players and the zone, in our post on X.
Training data
The SSVM learned from 4,062 Solo FNCS games from this season's qualifiers and cups. Every few seconds of every game, and at every zone call, we recorded a snapshot of each player still alive:
- The storm: the zone, the time until it moves and closes, players left alive, and the strength of the lobby.
- Position: the distance to the next circle and the speed needed to reach it, the terrain on the way in, and height, above the ground and against nearby players.
- Nearby players: how many are close by, how close the nearest player is, how many are in the next circle, and how many are in the way to the zone.
- Health and inventory: health, shield, materials, weapons (including rarity), heals and mobility items.
- Standing: eliminations and points so far, net damage, and storm surge standing.
- Recent play: damage dealt and taken in the last 30 seconds, time since landing, and loot and builds nearby.
Training the SSVM
We fit gradient-boosted decision trees to predict each player's final points from their snapshot, then tested the predictions on a test set of 727 Qualifier 2 Round 3 and Fast Track games that were excluded from training. In those games, players with a given number of expected points finished with about that many points on average, at landing, in the mid game, and in the late game.
Figure 1: Expected vs. Actual Final Points · Unseen Games
Expected points
Expected points are an average over the games that could follow from a player's situation. A player with 30 expected points won't score exactly 30, but players in that situation score 30 on average.
Zone luck
When the next zone is called, a player's expected points jump up or down depending on where it lands. To measure how lucky that call was, we ask the model what the player's expected points would have been under the other zones the game could have drawn, with everything else kept the same, and average them. Zone luck is the called zone's expected points minus that average.
Below is one player at a zone 6 call. Zone 6 moves a fixed distance in a random direction, so we sample circles evenly around the current zone (24 here). The called zone gave this player 31.0 expected points, against an average of 24.4, so the call was worth +6.6 points of zone luck.
Figure 2: Expected Points under 24 Possible Zone 6 Circles · One Player
Valuing items
The SSVM can also put a value on items. We can take a player's snapshot and add or remove one thing: an item, a stack of materials or a whole inventory. The change in expected points is what that thing is worth (at that moment of the game).
To get a value that holds in general, we repeat this for hundreds of thousands of real snapshots from many games and average the changes, separately for the early, mid and late game.