Ranked Player Distribution
Reading the lobby chart
Every ranked lobby in scope, bucketed by its average level. The highlighted bars are the slice your rank or skill setting is reading from — move the slider and the rest of the page recounts from those lobbies.
Ranked tier is the in-game ladder: 22 steps from Bronze I through Masters III to Pro. Skill is the model's own estimate, as a z-score — 0 is an average lobby, ±1 is a standard deviation. The range controls select the same lobbies on either chart. Skill is best compared across a full season, since ladder ranks shift as players climb.
Draft Agent
Backed by advanced machine learning and tree search strategies. Learn more
Your team
Your win chance
How likely your team is to win with the brawlers picked so far, on this map and at this rank. 50% is a coin flip.
It counts only picks already made, so it swings early and settles as the teams fill up.
More detail
The draft model is a logistic model of which team wins, trained on the season's ranked games by the draft agent's pipeline. Each team gets a score with three parts:
Standalone strength of each brawler: a base weight plus a shift for the map, the mode and the lobby's skill, w + ectx · (map + mode + skill × vskill). Synergy for every pair of team-mates: esyn · esyn. Matchups against every opponent: one side's attack vectors against the other's defence, minus the reverse, eatt · edef.
Win chance is the logistic function of your score minus theirs plus the matchup term. An empty slot adds nothing, so this is the board as it stands, not an average over the picks still to come; the pick percentages below are that average.
Opponents
Bans
A banned brawler is off the table for both teams, as in ranked above Mythic. The picks re-rank around it.
Tap ⊘ on any pick card to ban it from there.
More detail
A ban marks the brawler unavailable for the whole search: it cannot be a candidate, cannot be drawn for either team in any play-out, and never appears in the tree. The chances of every other later pick are renormalised over what is left, so banning a popular brawler or a common counter moves the numbers for every candidate that was weak against it or relied on it being gone. At least six brawlers must stay available.
Best picks for your team
Your chance to win if you pick it now.
Your rank
Set this to the rank you are drafting at. It changes how strong each brawler is rated and what the lobby is expected to pick. It only affects the draft agent.
Page range follows the rank range at the top of the page.
More detail
Rank reaches the draft in two ways. First, the model gives each lobby one skill value and moves every brawler's standalone strength along a direction it learned for skill, ectx · vskill × skill. Some brawlers gain with skill and some lose, which is enough to reorder close picks. Synergy and matchups are the same at every rank.
Second, the lobby's popularity (see Lobby) is counted from games at that rank, because what people pick changes with rank too.
A rank here means a whole tier: its skill value is the average over that tier's games on the selected map and dates. Tiers near the top are rare in ranked data, so if a tier has fewer than 20,000 of those games, popularity also counts the ranks just below it until it has enough, and a note says so. The skill value stays the tier's own. Page range uses the page's rank range for both, without changing it.
Lobby
To score a pick, the agent plays out the rest of the draft thousands of times. This sets how it imagines everyone else picking. It never changes how a board is scored.
| Typical | Players pick what is popular on this map at this rank, and lean toward counters as much as real ranked lobbies do. |
|---|---|
| Worst case | Every later pick is the best counter left. A stress test. |
More detail
In Typical, each later pick, on both teams, is drawn at random with chance proportional to popularity1.25 × ec × edge. Popularity is how often the brawler is picked in the games selected for the draft (map, dates and rank), with a small floor so a brawler nobody picks stays reachable. Edge is how much the brawler's matchups against the other team add to its team's win logit on that exact board. It is exact, not estimated, because every term of the model is linear in one added brawler.
Both constants are measured and refit every season. The exponent (1.25) is the one that best predicts held-out real picks. The counter strength c = 2 is found by comparing how often lopsided matchups turn up in real games with how often the same teams would meet at random, and matching that to drafts simulated at known strengths. Standalone strength and synergy get no extra weight: once popularity is known they add nothing to predicting real picks, because popularity already reflects them.
Worst case makes every later pick, for both teams, the one that adds the most to its team's win logit. Your own team's later picks are imagined the same way as the opponents' in both settings.
Search time
How long the agent keeps playing out the draft before it stops. A shorter time gives you a final answer sooner, but the numbers may not have settled yet, especially on a slower device. The picks can be clicked the whole time either way.
More detail
This is a cap on wall-clock time, in whole seconds, up to 14. The numbers settle from about 40,000 play-outs. A recent desktop browser gets there in one to two seconds; slower devices take longer. If the search stops before that, a note under the search tree says so, and picks within a point or two of each other should be read as ties.
Pick percentages
Your chance to win if you pick that brawler now, with the rest of the draft played out. On the opponents' turn, it is their chance.
The search runs for up to 14 seconds (see Search time) and the numbers settle as it goes. Low confidence means a pick has not been played out enough yet.
More detail
The agent runs a Monte Carlo search from the current board in a background worker. Every brawler still available is a candidate, and the search spreads its play-outs evenly over all of them, so every card is measured alike.
A play-out puts the candidate on the board, then fills the remaining picks in ranked order (one, two, two, one), drawing each from the Lobby policy. The final pick is not drawn: the win chance is averaged exactly over every brawler that could make it, weighted by its chance. A card's percentage is the mean over all of its play-outs, which makes it the expected win chance over the drafts that could follow that pick.
Results settle from about 40,000 play-outs in total. A card with fewer than 200 of its own says Low confidence, and Not searched is a one-pick estimate shown before the search reaches it. These numbers can sit above or below the win chance on the board, because they already allow for the picks still to come.
Inside the search
The live search behind the picks. Line weight is how often the search went that way; under each brawler is the average win chance there and how many play-outs passed through.
Point at a pick card, or click a brawler here, to open its branch.
More detail
This is the search's own tree, drawn from snapshots the worker sends while it runs, not an illustration. The root is the board as it stands; the first column is the eight leading candidates, and the open one shows the four replies that came up most in its play-outs, each followed by the likeliest line after it.
Below the root nothing is chosen: every later pick is drawn from the Lobby policy, so a reply's share of its parent's play-outs is how likely that reply is. Line width grows with the square root of that share (for candidates, of their play-outs relative to the most-searched one). Each node shows the mean win chance of every play-out through it and their count. The line under the open branch follows the most common reply at each step, and stops where a position has been reached fewer than three times.
The tree records up to 131,072 positions; past that, play-outs continue without being recorded, so the numbers keep improving. Replay slowly plays the recorded snapshots back at up to a third of real speed.
The Atlas
Reading the atlas
Color is strength: red is weaker and green is stronger. Size shows pick rate in the selected games. Drag to pan and use + / − to zoom. The numeric strength score is the model’s standalone score: higher is stronger, not a win probability.
Brawler positions come from learned embedding vectors built from all ranked games. Brawlers near each other play similarly; brawlers far apart play differently.
Positions stay fixed because the embedding vectors are based on all ranked games, not retrained for a map or Elo subset.
Drag to pan · Positions stay fixed across filters
Brawler Data Table
⚠ Fewer than 100 games in this slice; rates may be noisy. Strength is a model score, not a win probability.
| # | Brawler | Strength | Win rate | Picked | Games |
|---|