Skip to content

Search and nets

Search needs an engine that is cheap to copy and step. A net needs a runtime that evaluates it in pure Go. This page says what ships, and where each piece runs.

flowchart LR
  R[Root decision] --> W[Worlds]
  W --> T[Tree]
  N[Net] -.->|prior, value| T
  T --> A[Answer]

Which search runs where

Search gorge dirac
PUCT tree over redealt worlds (bots/azmcts) full none
Determinised rollouts (bots/searchseat) full none
Heuristic bot, no search full full (hbot)
Search primitives: copy, step, hide, redeal full full

dirac ships the primitives a search uses and its own bot. It ships no tree search, and gorged -engine dirac refuses a search policy. What dirac offers a search lists the operations.

Defaults from azmcts.DefaultOptions: 100 simulations per decision, exploration constant 1.5, first-play reduction 0.1, at most 6 candidates, 1,000 steps per simulation. Four decision kinds are searched: priority, attackers, blockers and target. FullKinds adds modes, may-triggers, small choices and multi-target picks. Every other decision goes to the default bot. Simulations have no rollouts: each leaf is scored once, by a heuristic or by a net.

The world source decides what the search may know.

fair redeal re-deals everything the seat can't see, from its own observations. redeal-closed also hides the opponent's deck list, drawing a believed list from a corpus you supply.

clairvoyant clones the real engine. The source for it is not in this repository, and the seat refuses it unless a command injects one.

az-redeal is the hosted entry: the redeal world, 100 simulations, no net, constructed or Commander, up to four seats, experimental tier. gorged.RegisterNetBot links a policy that loads a net at startup and adds -<policy>-net, -<policy>-decks, -<policy>-belief and related flags.

The determinised-rollout seat (bots/searchseat) samples K worlds per decision (default 8), rolls each out and overrides the bot only when a candidate wins by a margin. By default it covers attackers and the cast choice.

Per-simulation cost is on Bot speed.

The net runtime

Nets run in pure Go, with no cgo and no external ML runtime. Two packages read the same .dzgw file.

  • bots/gonet

    Loads mlp and gnn nets. Caches evaluations and card vectors. The graph net uses AVX-512 kernels when the CPU has them.

  • bots/netsimd

    Whole-layer AVX-512 kernels for the mlp net, with a scalar twin that gives the same bits. A library today; only its tests import it.

What Limit
Architectures mlp and gnn in gonet; mlp in netsimd
netsimd widths d, ff, card_d, head_hidden multiples of 32; d_emb any; 1 to 64 trunk blocks
Fast path d = head_hidden = 64
CPU paths AVX-512 (F, BW, DQ, VL) on amd64; scalar Go everywhere else, and for ForceScalar
Raw card width gonet 68 or 100; netsimd 68, 84 or 100

The file format and how a net is loaded are on Training data and nets.