Skip to content

Search and bots

Which searches and nets ship, and where they run: Search and nets.

dirac exists so that searches and self-play can step games cheaply. This page says what it offers a search, for bot authors and for readers comparing engines. It reports no playing strength. Speed per simulation is on Bot speed.

Need Operation Cost
copy a position dst.CopyFrom(root) no allocation (Performance)
list a decision StepInto(&d): packed Actions in canonical order no allocation
apply a move Apply(a) no allocation
hide what a seat can't see Hidden(o, seat), HiddenCards, DealHidden (re-deal hidden cards from a pool), RedealOwnLibrary, ReseedChance in place, on the copy
read a position ObjCard, ObjZone, Controller, HandSize, LibrarySize, StepNow and others no allocation
recompute only what changed SetEpoch and the change marks, ObjHash, ProjectDelta proportional to what changed

Everything here has to keep the rule the whole engine keeps: no allocation per transition, no map iteration whose order could reach an action, and the same seed and answers give the same game, byte for byte.

How a simulation uses them

The shape below is that of bots/azmcts, gorge's PUCT tree search over information sets. dirac/search is its port onto dirac, and dirac/neteval its network evaluator.

  1. Copy the root with CopyFrom.
  2. Re-deal what the seat can't see (its own library order, the opponent's hand and library, from a card pool if one is given) and re-seed chance, from a per-decision seed.
  3. Walk: play the tree's edges as packed actions, and answer every other decision with a heuristic bot until the next searched point.
  4. Expand one node, value the leaf once with an evaluator (no rollouts), and back the value up.

Almost all of a simulation is the engine stepping along that line, which is why the transition and copy numbers decide what a search can afford. A rules probe inside Step uses a fork from the State's fork pool, so nothing allocates at a warmed depth.

The ports

Package Entry points What it gives a bot
dirac/search DefaultOptions(), NewSearcher(opts, eval, bot), (*Searcher).Search, (*Searcher).SeatSearch the PUCT search above, its candidates, re-deals and option families
dirac/neteval LoadWeights(path), New(w, pool) a network search.Evaluator: dirac/feat feeding bots/netsimd or bots/gonet
dirac/search/hbot NewFacts(*kop.Table), New(*Facts), (*Bot).EnableHoldInstant() a native port of gorge's auto-pay heuristic bot, held to it decision by decision by khbot
dirac/search/mull KeepLands, KeepConstructed, KeepNever, CoinMulls gorge's London-mulligan policies, as pure functions over a hand
dirac/feat NewPool(tbl, specs), NewEncoder(pool), EncodeState, EncodeOptions the network input encoding, below

These import nothing from gorge, except dirac/neteval, which imports gorge's network packages (bots/netsimd, bots/gonet, bots/policynet) and never the engine. oracle.HBotFacts and oracle.FeatPool build their tables from a gorge registry, and dirac-only code passes a kop.Table and pre-derived specs instead.

Network inputs

dirac/feat turns a position into the same inputs gorge's bots/policynet produces from the deciding seat's redacted view: dense scalars, a sorted hashed sparse bag, one entity record per visible card, and one encoding per option. Lockstep can assert that byte for byte at every decision, so a network trained on one engine's games reads the other's identically.

dirac has no strings, so every string-derived fact the encoder reads (names, type lines, costs, keyword spellings) is computed per CardID into a Pool when the pool loads. The hot path streams FNV-1a over those rows and dirac's own numbers, and an encode allocates nothing.