War decks

Four decks, thirty-two cards, chosen exactly from what this season's players played, and what to upgrade to make them better.

Deck sets

Clan Wars asks a member for four decks with 32 distinct cards, and a card and its Evolution or Hero form are the same card (the tower troop is not one of the 32). battles_deck_sets answers "which four should I play" from the record rather than a guess: it reads the decks this season's recorded players played, keeps the ones the player can field, values each, and packs the best sets exactly.

A deck here is its eight cards. Clan Wars battles carry no tower troop (the API sends none on a war battle), while Trophy Road and Path of Legends battles do, so the same eight cards are several deck_hash values in the record: every one's record pools into the deck's (variants lists them, most played first, and deck_hash is the first). The player's own Clan Wars duel rounds count too: a duel is one battle with up to three games, each game its own deck, so each round is a game on its eight cards, won or lost by that round's crowns (your_duel_rounds, and modes.war.your_duel_rounds). A war deck the player plays only in duels is a candidate like any other.

The reductions, in order:

Each remaining deck is valued in log-odds, and every part rides the row (value):

• corpus_logit: its record over Trophy Road, Path of Legends and Clan Wars (modes carries each). Each mode's rate is pulled toward that mode's corpus mean as the meta tools' rates are (the formula is on the methodology page; priors carries the means), and a Trophy Road or Clan Wars rate is corrected for its players' level edge (mean_level_gap) at 0.5 log-odds per level, the effect measured on the corpus in both modes. Path of Legends equalises levels, so its rows are not corrected. The modes are pooled by battles.
• level_term: 0.5 per level the player would field the deck above or below their target level.
• form_term: for each card played as a form the player has not unlocked, minus that card's measured form advantage this season (its form's rate against its base form's over every mode, in log-odds, never a bonus); a card whose forms are too thin to measure takes the season's median (measured: false).
• familiarity_term: 0.05 when the player has played these eight cards five or more times this season, a tie-break for a deck they know.

A set's value is its decks' values plus the weakest deck's again, locked decks included, so a set is never carried by three strong decks and one weak one: every war day asks for all four. The search is exact (branch and bound over the best 1,500 candidates); search.exhausted false says it stopped at its budget. When the defaults find no set, the candidates are widened once (min_battles 5, min_players 2, no level floor: the gap is priced by level_term and each deck's fit.lowest_card shows its weakest card) and applied says so. When no whole set exists even then, partial_set is the best set of fewer decks that does, so the answer is never a dead end. alternatives returns more sets, each differing from every earlier one in at least two of the decks the search chose, and near_misses names decks worth at least the first set's weakest that it gave up, with the cards they share with a chosen deck.

Shape the set in conversation: lock_decks keeps decks (a deck_hash from battles_decks, battles_meta_decks or an earlier answer; any variant names the same eight cards) and fills the rest, so "a different last war deck" is the other three locked and the current fourth in exclude_decks, which no set may choose; exclude_cards keeps cards out; require_cards puts cards in. The arguments are settled before any search: a card both required and excluded, a required card the player does not hold, and locked decks that share cards are refused, naming the cards. locked_decks echoes every locked deck with its fit, so a locked deck the player cannot field says so. A value is an ordering, not a forecast: a deck's record is its players', and pilots differ.

Deck upgrades

battles_deck_upgrades is the other half of a war set: not "which four can I play" but "what should I upgrade so my four get better". It starts from the player's best set today (the same card sets, values and exact packing as battles_deck_sets, over the decks the player owns every card of, with no level floor: the gap is what an upgrade closes) and prices single upgrades:

• a card raised toward the level the player fields (to_level, at most max_levels at a time), for every card below that level in a deck that could join the set;
• an Evolution or Hero form unlocked that a candidate deck plays, which takes that card's measured form price off every deck that uses it.

For each option the set is re-packed exactly and gain is its value after minus before, in log-odds: the change in the same set value battles_deck_sets optimises, never a score of its own. lifts names the decks of the set whose value it raises, set_changes says whether the upgrade changes which decks the set holds, and set_after shows the set it gives. Options are priced one at a time and do not add up: take the first, then ask again. Levels, not gold: the game's upgrade costs are not in the record, and a card the player does not own is never an option.

One card raised rarely moves a deck whose every card sits two levels under, so within_reach lists the decks outside the set that would join it once their low cards reach the fielded level: each card at most max_levels under, raises naming every card and its levels, forms any Evolution or Hero form it plays that the player has not unlocked, levels the total, and the same gain and set_after.