j140 rename the lead-gain corrector: BitBrain -> LEADGAIN (+ legacy TR_BITBRAIN_* aliases)
The gun at rack id 16 learned a multiplier for Pattern's lead, separately
per range band. It was called BITBRAIN and shipped a TR_BITBRAIN_* prefix,
which is why the name read as a neural network it no longer contains.
guns/bitbrain_gun.nim -> guns/lead_gain.nim (rack id 16 UNCHANGED)
RackGunNames[16] BITBRAIN -> LEADGAIN
TR_BITBRAIN_* knobs -> TR_LEADGAIN_*
[bb] log line -> [lg]
BACKWARD COMPATIBILITY is mandatory: the live .env carries
TR_RACK_BITBRAIN=both, TR_BITBRAIN_GAINS, TR_BITBRAIN_MEM=decay and
TR_BITBRAIN_LOG=1, and those must keep behaving identically. The new ADE+SBC
gun (next commit) claims the BITBRAIN name and the TR_BITBRAIN_* prefix, so
the namespace is disambiguated by ONE deterministic switch, TR_BITBRAIN_NET
(default 0):
TR_BITBRAIN_NET unset/0 -> LEGACY: the 14 frozen legacy suffixes are aliases
for TR_LEADGAIN_*, and TR_RACK_BITBRAIN still
selects rack id 16. One [depr] line on stderr
names the new spelling of each honoured knob.
TR_BITBRAIN_NET = 1 -> the TR_BITBRAIN_* names belong to the new gun.
The legacy suffix set and the new gun's knob set are DISJOINT, so no name is
ever claimed twice; the new name always wins over its alias.
Parity: shipped rack is still onlyPattern, shipped movement is still strafe.
Guards unchanged: test_env_report 25, test_rack_membership 48,
test_tm_pattern_registration 20, test_lead_gain_registration 13 (was
test_bitbrain_registration), test_bitbrain 56, test_gun_harness 39,
test_tfil_commit_env 30. New: test_lead_gain_legacy 24.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
@@ -0,0 +1,616 @@
|
||||
## lead_gain.nim — LEADGAIN (rack id 16): a per-range-band LEAD-GAIN corrector.
|
||||
##
|
||||
## ── THE NAME ──────────────────────────────────────────────────────────────────
|
||||
## This gun used to be called `BITBRAIN` and to live in `guns/bitbrain_gun.nim`,
|
||||
## but its ADE+SBC network was removed when it was rebuilt into what it actually
|
||||
## is: **it learns a multiplier for Pattern's lead, separately per range band.**
|
||||
## `LEADGAIN` says that; `BITBRAIN` (a neural network) did not. The rack id 16
|
||||
## is UNCHANGED (many tests assert the id literals) and the real ADE+SBC gun is
|
||||
## the separate `guns/bitbrain_net.nim` at rack id 17.
|
||||
##
|
||||
## ── WHY THE FILE WAS REBUILT (Phase 0/1 evidence) ────────────────────────────
|
||||
## The previous design was an ADDITIVE angular shift: an ADE+SBC network
|
||||
## classified the +h-tick angular error over ±`TR_LEADGAIN_RANGE` degrees and
|
||||
## added the argmax class centre to Pattern's bearing. Phase 0 measured it as
|
||||
## statistically identical to Pattern (`docs/bitbrain_gun_verdict.md`,
|
||||
## commit d93ce44) and as carrying no measurable aim information
|
||||
## (450+: 16.200 deg vs Pattern's 16.193; `docs/bitbrain_campaign.md` §0.3.5).
|
||||
##
|
||||
## Phase 1 measured the actual lever. The gain sweep found that gains >= 1 are
|
||||
## strictly worse at every band and that the optimal gain is BELOW 1.0 at long
|
||||
## range (450+: ~0.25). A fractional gain leaves the Pearson lead *correlation*
|
||||
## unchanged (correlation is invariant under positive scaling), so a smaller
|
||||
## gain does not add information — it shrinks the magnitude of an uninformative
|
||||
## Pattern lead toward the low-variance static (HeadOn) aim. The right output is
|
||||
## therefore a multiplicative GAIN on Pattern's lead, not a class-based additive
|
||||
## shift. See `docs/bitbrain_campaign.md` §Phase 1 for the measured curve.
|
||||
##
|
||||
## ── THE DESIGN ────────────────────────────────────────────────────────────────
|
||||
## * BASE — the shipped Pattern gun's prediction (`guns/pattern_matcher`),
|
||||
## reached through the TmHorizonGun observation ring.
|
||||
## * OUTPUT — `aim = LOS + gain * (patternAim - LOS)`, i.e. Pattern's lead over
|
||||
## the line of sight is multiplied by a learned `gain` (one of
|
||||
## `LG_CAND`, so it may be BELOW 1.0 — the point).
|
||||
## * LABEL — the same deferred-label path the old corrector used: at fire
|
||||
## time we remember the base lead and the aim tolerance; `h =
|
||||
## round(dist/speed)` ticks later `tmhObservedAt` returns the
|
||||
## enemy's OBSERVED bearing from the firing position. `requiredLead
|
||||
## = observedBearing - LOS` and `baseLead = baseBearing - LOS`, so a
|
||||
## candidate gain scores a hit on this sample when
|
||||
## `|gain*baseLead - requiredLead| <= tolerance`.
|
||||
## * TRAIN — ONLINE / PREQUENTIAL per range band: for each candidate gain we
|
||||
## count the fraction of resolved samples that would have been
|
||||
## within the target's angular half-width (`atan(18/range)`, the
|
||||
## SAME tolerance the offline ruler uses). The band's gain is the
|
||||
## argmax hit rate. THIS is the key lesson of Phase 1: the
|
||||
## least-squares gain and the hit-probability-optimal gain DIVERGE
|
||||
## (Pattern's lead errors are bimodal), so the learner optimises the
|
||||
## hit-probability proxy directly instead of mean squared error.
|
||||
## * STATE — the range band (the ruler's 5 bands). Range is known causally at
|
||||
## fire time, so a per-band gain table is shippable with no learning
|
||||
## at all; this gun learns that table online. The correction is
|
||||
## additionally gated to bands with range >= 300 px
|
||||
## (`LG_GAIN_BAND_MIN`), where Phase 1 measured Pattern's lead to be
|
||||
## uninformative. That gate is causal (range is known).
|
||||
##
|
||||
## The gain statistics are battle-scale: a round boundary wipes the observation
|
||||
## ring and deferred labels but NOT the gain counts (a new round is not a new
|
||||
## enemy). `resetLearning` wipes them on a new battle / target change; with
|
||||
## `TR_LEADGAIN_MEM=decay` every `TR_LEADGAIN_DECAY` resolved samples decays the
|
||||
## counts by `TR_LEADGAIN_DECAY_FRAC` toward the gain-1.0 column.
|
||||
##
|
||||
## ── WHAT IS STILL HERE ONLY FOR THE BOOT REPORT / GUARD TESTS ─────────────────
|
||||
## The ADE+SBC network is GONE from the gun. The class geometry
|
||||
## (`lgCenterDeg`/`lgClassOf`) and `TR_LEADGAIN_N`/`NADE`/`WARMUP`/`ADAPT`/
|
||||
## `CALIB`/`SEED`/`RANGE` are retained as resolved configuration so the boot
|
||||
## report (`env_report.nim`) and the registration guard tests keep working
|
||||
## unchanged; they no longer affect the gain learner. The generic
|
||||
## `common_libs/bitbrain/` library is untouched and still tested by
|
||||
## `test_bitbrain.nim`, and the new ADE+SBC gun that actually uses it is
|
||||
## `guns/bitbrain_net.nim` (rack id 17).
|
||||
##
|
||||
## ── BACKWARD COMPATIBILITY: the `TR_BITBRAIN_*` legacy aliases ───────────────
|
||||
## The owner's live `.env` predates the rename and contains `TR_RACK_BITBRAIN`,
|
||||
## `TR_BITBRAIN_GAINS`, `TR_BITBRAIN_MEM`, `TR_BITBRAIN_LOG`, … Those names are
|
||||
## the OLD corrector's knobs and MUST keep working unchanged. The
|
||||
## `TR_BITBRAIN_*` prefix, however, now belongs to the NEW ADE+SBC gun
|
||||
## (`guns/bitbrain_net.nim`). The two uses are separated by ONE deterministic
|
||||
## switch, `TR_BITBRAIN_NET` (the new gun's master switch, default 0 = off):
|
||||
##
|
||||
## * `TR_BITBRAIN_NET` UNSET / 0 → LEGACY MODE. Every `TR_BITBRAIN_<X>` name
|
||||
## listed in `LegacyKnobEnvNames` is a legacy alias for this gun's
|
||||
## `TR_LEADGAIN_<X>`, and the new ADE+SBC gun is OFF. This is the owner's
|
||||
## current configuration, so its behaviour is unchanged.
|
||||
## * `TR_BITBRAIN_NET=1` → NEW-NETWORK MODE. `TR_BITBRAIN_<X>` names
|
||||
## the NEW gun's knobs (see `bitbrain_net.nim`) and this gun reads ONLY
|
||||
## `TR_LEADGAIN_<X>`.
|
||||
##
|
||||
## The legacy sets are DISJOINT (see `legacyKnobEnvNames` / the new gun's
|
||||
## `netKnobEnvNames`), so no name is ever claimed by both. `TR_RACK_BITBRAIN` is
|
||||
## the one genuinely ambiguous name (the rack is keyed by gun name and the new
|
||||
## gun is now the one called `BITBRAIN`); it is resolved by the SAME switch —
|
||||
## see `selector.nim`'s `RackLegacyAliases`.
|
||||
##
|
||||
## ── TR_LEADGAIN_GAINS (the candidate set as an env knob) ─────────────────────
|
||||
## `TR_LEADGAIN_GAINS` is a comma-separated candidate list, e.g.
|
||||
## `TR_LEADGAIN_GAINS=1.0,1.25,1.5,2.0`. It replaces the fixed shipped candidate
|
||||
## set `{0, 0.25, 0.5, 0.75, 1.0}` for this gun instance, so every live arm is
|
||||
## pure-env (no recompile). Two degenerate cases are deliberate:
|
||||
## * unset / unparsable -> the shipped `LG_CAND` set, byte-identical behaviour;
|
||||
## * exactly ONE value -> a FIXED gain, applied from the first shot with NO
|
||||
## learning at all (the learner is bypassed), still gated to the long bands.
|
||||
## The applied `gain` (and the resulting angular `shift`) is printed on the
|
||||
## existing change-gated `[lg]` line, so a run's liveness AND the correction it
|
||||
## actually applied are both auditable from the bot's stdout.
|
||||
##
|
||||
## DEFAULT OFF / PARITY: this gun is admitted ONLY when `TR_RACK_LEADGAIN` says so
|
||||
## (default `off`). The shipped rack never calls `predict`, so `ensureInit` never
|
||||
## runs and the shipped bot is byte-for-byte unchanged.
|
||||
|
||||
import std/[math, os, strutils, strformat, algorithm]
|
||||
import gun_harness/gun_interface
|
||||
import guns/tm_horizon
|
||||
import guns/pattern_matcher
|
||||
|
||||
const
|
||||
## ── env knobs (all resolved once at gun construction) ─────────────────────
|
||||
LG_MEM_ENV* = "TR_LEADGAIN_MEM" ## perRound|retained|decay
|
||||
LG_GAINS_ENV* = "TR_LEADGAIN_GAINS" ## comma-separated candidate gains
|
||||
LG_N_ENV* = "TR_LEADGAIN_N" ## (legacy geometry; inert)
|
||||
LG_NADE_ENV* = "TR_LEADGAIN_NADE" ## (legacy ADE count; inert)
|
||||
LG_RANGE_ENV* = "TR_LEADGAIN_RANGE" ## (legacy class half-range; inert)
|
||||
LG_LOG_ENV* = "TR_LEADGAIN_LOG" ## 1 = per-change [lg] log
|
||||
## ── the one switch that disambiguates the legacy `TR_BITBRAIN_*` names ────
|
||||
## Read by BOTH guns (see `bitbrain_net.nim`). Unset/0 => the `TR_BITBRAIN_*`
|
||||
## names are LEGACY aliases for this gun; 1 => they belong to the new ADE+SBC
|
||||
## gun. It is also the new gun's master on/off switch.
|
||||
LG_NET_SWITCH_ENV* = "TR_BITBRAIN_NET"
|
||||
LG_MIN_OBS_ENV* = "TR_LEADGAIN_MIN_OBS" ## samples before a band is trusted
|
||||
LG_WARMUP_ENV* = "TR_LEADGAIN_WARMUP" ## (legacy; inert)
|
||||
LG_ADAPT_ENV* = "TR_LEADGAIN_ADAPT" ## (legacy; inert)
|
||||
LG_CALIB_ENV* = "TR_LEADGAIN_CALIB" ## (legacy; inert)
|
||||
LG_DECAY_ENV* = "TR_LEADGAIN_DECAY" ## decay interval (samples)
|
||||
LG_DECAY_FRAC_ENV* = "TR_LEADGAIN_DECAY_FRAC" ## per-decay count shrink
|
||||
LG_SEED_ENV* = "TR_LEADGAIN_SEED" ## (legacy; inert)
|
||||
LG_RESET_ON_TARGET_ENV* = "TR_LEADGAIN_RESET_ON_TARGET"
|
||||
## ── fixed geometry ────────────────────────────────────────────────────────
|
||||
LG_PENDING_CAP* = 512 ## deferred-label queue (>= 4 buckets x 50 ticks)
|
||||
## ── the gain learner ──────────────────────────────────────────────────────
|
||||
LG_NBANDS* = 5 ## the ruler's range bands
|
||||
LG_NHB* = 4 ## horizon buckets (for the per-tick label dedupe)
|
||||
LG_BAND_LO* = [0.0, 100.0, 200.0, 300.0, 450.0]
|
||||
LG_BAND_HI* = [100.0, 200.0, 300.0, 450.0, 1.0e18]
|
||||
## The DEFAULT candidate lead gains the band selector picks from. 0.0 == HeadOn
|
||||
## (aim at the current position) and 1.0 == Pattern (use the full lead).
|
||||
## `TR_LEADGAIN_GAINS` replaces this set per gun; unset -> this exact set.
|
||||
LG_CAND* = [0.0, 0.25, 0.50, 0.75, 1.0]
|
||||
LG_NCAND* = 5
|
||||
LG_BOT_RADIUS* = 18.0 ## hit-detection radius in px (ruler tolerance)
|
||||
## Apply the correction only from this band up (range >= LG_BAND_LO[3] = 300).
|
||||
## [MEASURED] below 300 Pattern's lead is informative and shrinking it loses
|
||||
## hits; see the header note.
|
||||
LG_GAIN_BAND_MIN* = 3
|
||||
## ── shipped defaults ──────────────────────────────────────────────────────
|
||||
LG_N_DEF = 32
|
||||
LG_NADE_DEF = 256
|
||||
LG_RANGE_DEF = 40.0
|
||||
LG_MIN_OBS_DEF = 8
|
||||
LG_WARMUP_DEF = 400
|
||||
LG_ADAPT_DEF = 32
|
||||
LG_CALIB_DEF = 512
|
||||
LG_DECAY_DEF = 250
|
||||
LG_DECAY_FRAC_DEF = 0.02
|
||||
LG_SEED_DEF = 20240921
|
||||
LG_RESET_ON_TARGET_DEF = true
|
||||
|
||||
type
|
||||
LeadMemMode* = enum
|
||||
lgPerRound, lgRetained, lgDecay
|
||||
|
||||
LgPending = object
|
||||
## One deferred training sample. `lead` is Pattern's lead over LOS at fire
|
||||
## time (radians) and `tol` the target's angular half-width then; the label
|
||||
## is resolved `horizon` ticks later.
|
||||
fireTick: int
|
||||
horizon: int
|
||||
band: int
|
||||
selfX*, selfY: float
|
||||
baseBearing: float
|
||||
lead: float
|
||||
tolDeg: float
|
||||
|
||||
LeadGainGun* = object
|
||||
tmh: TmHorizonGun
|
||||
initialized: bool
|
||||
# ── resolved config (kept in the boot report) ─────────────────────────────
|
||||
nClasses*: int
|
||||
maxDeg*: float
|
||||
nAde*: int
|
||||
memMode*: LeadMemMode
|
||||
logEnabled*: bool
|
||||
minObs*: int
|
||||
warmupN*: int
|
||||
adaptEvery*: int
|
||||
calibEvery*: int
|
||||
decayEvery*: int
|
||||
decayFrac*: float
|
||||
seed*: int64
|
||||
resetOnTarget*: bool
|
||||
# ── the candidate gain set (resolved once at construction) ────────────────
|
||||
## Ascending; one entry == a FIXED gain with no learning. `bandHits` and
|
||||
## `bandN` are sized to it, so the loops below never touch a stale column.
|
||||
cands*: seq[float]
|
||||
# ── gain learner: hit counts per (range band x candidate gain) ────────────
|
||||
bandHits*: seq[seq[float64]]
|
||||
bandN*: seq[float64]
|
||||
trained*: int
|
||||
sinceDecay: int
|
||||
decays*: int
|
||||
# ── readout / accounting ──────────────────────────────────────────────────
|
||||
lastGain*: array[LG_NBANDS, float]
|
||||
corrections*: int
|
||||
lastLogKey: string
|
||||
# ── deferred labels ───────────────────────────────────────────────────────
|
||||
pending: array[LG_PENDING_CAP, LgPending]
|
||||
pendingCount*: int
|
||||
pendingDropped*: int
|
||||
# ── per-tick caches ───────────────────────────────────────────────────────
|
||||
lastTick: int
|
||||
lastEnqTick: int
|
||||
lastEnqBucket: int
|
||||
observedTargetId*: int
|
||||
|
||||
# ── small pure helpers ───────────────────────────────────────────────────────
|
||||
|
||||
proc wrapRadLg(r: float): float {.inline.} =
|
||||
result = r
|
||||
while result > PI: result -= 2.0 * PI
|
||||
while result < -PI: result += 2.0 * PI
|
||||
|
||||
proc memModeName*(m: LeadMemMode): string =
|
||||
case m
|
||||
of lgPerRound: "perRound"
|
||||
of lgRetained: "retained"
|
||||
of lgDecay: "decay"
|
||||
|
||||
proc lgGainsString*(cands: seq[float]): string =
|
||||
## The resolved candidate set as the env's comma-separated form (boot report).
|
||||
for i, c in cands:
|
||||
if i > 0: result.add ","
|
||||
result.add $c
|
||||
|
||||
proc parseLgGains*(value: string): seq[float] =
|
||||
## Parse `TR_LEADGAIN_GAINS`. Empty / unparsable / out-of-range / duplicate
|
||||
## input cannot silently select a different regime: it falls back to the
|
||||
## shipped `LG_CAND` set, exactly like the other env knobs fall back to their
|
||||
## defaults. Values are clamped to [0, 8] (0 == HeadOn, 1 == Pattern) and
|
||||
## de-duplicated, then sorted so the argmax tie rule (keep the smaller
|
||||
## candidate) is unchanged.
|
||||
var seen: seq[float]
|
||||
for tok in value.split(','):
|
||||
let t = tok.strip()
|
||||
if t.len == 0: continue
|
||||
var v: float
|
||||
try: v = parseFloat(t)
|
||||
except ValueError: continue
|
||||
if v < 0.0 or v > 8.0: continue
|
||||
var dup = false
|
||||
for u in seen:
|
||||
if abs(u - v) < 1e-9: dup = true
|
||||
if not dup: seen.add v
|
||||
if seen.len == 0:
|
||||
for c in LG_CAND: seen.add c
|
||||
return seen
|
||||
seen.sort()
|
||||
seen
|
||||
|
||||
proc parseLgMemMode*(value: string): LeadMemMode =
|
||||
## Empty / unknown values fall back to the shipped `perRound`, so a typo
|
||||
## cannot silently select another regime.
|
||||
case value.strip().toLowerAscii()
|
||||
of "retained", "retain", "accum", "accumulate": lgRetained
|
||||
of "decay", "forget", "age": lgDecay
|
||||
else: lgPerRound
|
||||
|
||||
proc lgEnv(name: string): string
|
||||
## Forward declaration: the legacy-alias lookup is defined below, after the
|
||||
## frozen `LegacyKnobEnvNames` table it depends on.
|
||||
|
||||
proc envFloatLg(name: string, default: float): float =
|
||||
let v = lgEnv(name)
|
||||
if v.len == 0: return default
|
||||
try: parseFloat(v.strip()) except ValueError: default
|
||||
|
||||
# ── legacy `TR_BITBRAIN_*` aliases (backward compatibility) ───────────────────
|
||||
|
||||
const
|
||||
LegacyPrefix* = "TR_BITBRAIN_"
|
||||
NewPrefix* = "TR_LEADGAIN_"
|
||||
## The COMPLETE, FROZEN set of the old corrector's knob suffixes. A
|
||||
## `TR_BITBRAIN_<X>` in this set is a legacy alias for `TR_LEADGAIN_<X>`; any
|
||||
## other `TR_BITBRAIN_*` name belongs to the new ADE+SBC gun
|
||||
## (`bitbrain_net.nim`). The two sets are DISJOINT by construction, so the
|
||||
## mapping is total and deterministic — no name is claimed twice.
|
||||
LegacyKnobEnvNames* = [
|
||||
"GAINS", "MEM", "MIN_OBS", "DECAY", "DECAY_FRAC", "LOG", "RESET_ON_TARGET",
|
||||
"N", "NADE", "RANGE", "WARMUP", "ADAPT", "CALIB", "SEED"]
|
||||
## Knobs that actually change behaviour (the rest are inert configuration kept
|
||||
## for the boot report). A deprecation line is only worth printing for these
|
||||
## plus the inert ones, because a stale inert name is still a stale name.
|
||||
LegacyRackEnvName* = "TR_RACK_BITBRAIN"
|
||||
|
||||
proc netSwitchOn*(): bool =
|
||||
## `TR_BITBRAIN_NET` unset/0 => the `TR_BITBRAIN_*` names are LEGACY aliases
|
||||
## for this gun. 1 => they belong to the new ADE+SBC gun. The same predicate
|
||||
## is defined in `gun_harness/selector` (`netSwitchOwnsBitbrainName`), which
|
||||
## cannot import a concrete gun module.
|
||||
case getEnv(LG_NET_SWITCH_ENV, "").strip().toLowerAscii()
|
||||
of "1", "true", "yes", "on": true
|
||||
else: false
|
||||
|
||||
var deprecationShown = false
|
||||
|
||||
proc lgDeprecationLine*(): string =
|
||||
## The single clear deprecation line the owner sees. Names every legacy
|
||||
## `TR_BITBRAIN_*` knob that is actually set in the environment and the new
|
||||
## name that now owns it. Empty when there is nothing to migrate.
|
||||
if netSwitchOn(): return ""
|
||||
var parts: seq[string]
|
||||
for suffix in LegacyKnobEnvNames:
|
||||
let old = LegacyPrefix & suffix
|
||||
if getEnv(old, "").len > 0:
|
||||
parts.add old & " -> " & NewPrefix & suffix
|
||||
if getEnv(LegacyRackEnvName, "").len > 0:
|
||||
parts.add LegacyRackEnvName & " -> TR_RACK_LEADGAIN"
|
||||
if parts.len == 0: return ""
|
||||
result = "[depr] " & LegacyPrefix & "* is the OLD lead-gain corrector's namespace; " &
|
||||
"it was renamed to " & NewPrefix & "* (gun LEADGAIN, rack id 16). " &
|
||||
"Still honoured: " & parts.join("; ") &
|
||||
". The new ADE+SBC gun owns the " & LegacyPrefix &
|
||||
"* names once " & LG_NET_SWITCH_ENV & "=1."
|
||||
|
||||
proc lgEnv(name: string): string =
|
||||
## Read a `TR_LEADGAIN_<X>` knob, falling back to the legacy
|
||||
## `TR_BITBRAIN_<X>` alias while `TR_BITBRAIN_NET` is off. The NEW name always
|
||||
## wins when both are set, so a migrated config is authoritative.
|
||||
var v = getEnv(name, "")
|
||||
if v.len > 0: return v
|
||||
if netSwitchOn(): return ""
|
||||
let suffix = if name.startsWith(NewPrefix): name[NewPrefix.len .. ^1] else: ""
|
||||
if suffix.len == 0: return ""
|
||||
for s in LegacyKnobEnvNames:
|
||||
if s == suffix: return getEnv(LegacyPrefix & suffix, "")
|
||||
""
|
||||
|
||||
proc envIntLg(name: string, default: int): int =
|
||||
let v = lgEnv(name)
|
||||
if v.len == 0: return default
|
||||
try: parseInt(v.strip()) except ValueError: default
|
||||
|
||||
proc envBoolLg(name: string, default: bool): bool =
|
||||
case lgEnv(name).strip().toLowerAscii()
|
||||
of "1", "true", "yes", "on": true
|
||||
of "0", "false", "no", "off": false
|
||||
else: default
|
||||
|
||||
proc lgCenterDeg*(k, nClasses: int, maxDeg: float): float =
|
||||
## Centre (degrees) of correction class `k` over ±maxDeg. Retained for the
|
||||
## registration guard test and the boot report; inert for the gain learner.
|
||||
let w = 2.0 * maxDeg / float(nClasses)
|
||||
-maxDeg + (float(k) + 0.5) * w
|
||||
|
||||
proc lgClassOf*(errRad: float, nClasses: int, maxDeg: float): int =
|
||||
## Bin a signed angular error (radians) into one of `nClasses` bins over
|
||||
## [−maxDeg, +maxDeg]. Retained for the registration guard test; inert.
|
||||
let x = radToDeg(errRad)
|
||||
var k = int((x + maxDeg) / (2.0 * maxDeg) * float(nClasses))
|
||||
if k < 0: k = 0
|
||||
if k >= nClasses: k = nClasses - 1
|
||||
k
|
||||
|
||||
proc lgBandOf*(range: float): int {.inline.} =
|
||||
## Range band (the ruler's bands), known causally at fire time.
|
||||
for b in 0 ..< LG_NBANDS:
|
||||
if range >= LG_BAND_LO[b] and range < LG_BAND_HI[b]: return b
|
||||
LG_NBANDS - 1
|
||||
|
||||
proc lgTolDeg*(range: float): float {.inline.} =
|
||||
## The target's angular half-width at `range` — atan(18/range) — i.e. the exact
|
||||
## tolerance the offline ruler uses for its hit-probability proxy.
|
||||
radToDeg(arctan2(LG_BOT_RADIUS, max(range, 1e-9)))
|
||||
|
||||
# ── construction / lazy init ─────────────────────────────────────────────────
|
||||
|
||||
proc initLeadGainGun*(): LeadGainGun =
|
||||
result.nClasses = clamp(envIntLg(LG_N_ENV, LG_N_DEF), 2, 512)
|
||||
result.nAde = clamp(envIntLg(LG_NADE_ENV, LG_NADE_DEF), 8, 4096)
|
||||
result.maxDeg = clamp(envFloatLg(LG_RANGE_ENV, LG_RANGE_DEF), 1.0, 180.0)
|
||||
result.memMode = parseLgMemMode(lgEnv(LG_MEM_ENV))
|
||||
result.logEnabled = envBoolLg(LG_LOG_ENV, false)
|
||||
result.minObs = max(1, envIntLg(LG_MIN_OBS_ENV, LG_MIN_OBS_DEF))
|
||||
result.warmupN = max(0, envIntLg(LG_WARMUP_ENV, LG_WARMUP_DEF))
|
||||
result.adaptEvery = max(1, envIntLg(LG_ADAPT_ENV, LG_ADAPT_DEF))
|
||||
result.calibEvery = max(1, envIntLg(LG_CALIB_ENV, LG_CALIB_DEF))
|
||||
result.decayEvery = max(1, envIntLg(LG_DECAY_ENV, LG_DECAY_DEF))
|
||||
result.decayFrac = clamp(envFloatLg(LG_DECAY_FRAC_ENV, LG_DECAY_FRAC_DEF), 0.0, 1.0)
|
||||
result.seed = int64(envIntLg(LG_SEED_ENV, LG_SEED_DEF))
|
||||
result.resetOnTarget = envBoolLg(LG_RESET_ON_TARGET_ENV, LG_RESET_ON_TARGET_DEF)
|
||||
result.cands = parseLgGains(lgEnv(LG_GAINS_ENV))
|
||||
result.bandN = newSeq[float64](LG_NBANDS)
|
||||
result.bandHits = newSeq[seq[float64]](LG_NBANDS)
|
||||
for b in 0 ..< LG_NBANDS:
|
||||
result.bandHits[b] = newSeq[float64](result.cands.len)
|
||||
result.lastTick = -1
|
||||
result.lastEnqTick = -1
|
||||
result.lastEnqBucket = -1
|
||||
result.observedTargetId = -1
|
||||
for b in 0 ..< LG_NBANDS: result.lastGain[b] = 1.0
|
||||
# ONE deprecation line per process, naming the new `TR_LEADGAIN_*` names.
|
||||
let dep = lgDeprecationLine()
|
||||
if dep.len > 0 and not deprecationShown:
|
||||
deprecationShown = true
|
||||
stderr.writeLine(dep)
|
||||
|
||||
proc ensureInit*(g: var LeadGainGun) =
|
||||
## Build the observation ring on first use. No network, no global-RNG use, so
|
||||
## the shipped default path is untouched and construction stays cheap.
|
||||
if g.initialized: return
|
||||
g.initialized = true
|
||||
g.tmh = initTmHorizonGun()
|
||||
|
||||
# ── the gain learner ─────────────────────────────────────────────────────────
|
||||
|
||||
proc lgAccumulate(g: var LeadGainGun, leadDeg, reqDeg, tolDeg: float, band: int) =
|
||||
## Score every candidate gain on this resolved sample: a candidate "hits" when
|
||||
## it would have put the aim within the target's angular half-width.
|
||||
for ci in 0 ..< g.cands.len:
|
||||
if abs(g.cands[ci] * leadDeg - reqDeg) <= tolDeg:
|
||||
g.bandHits[band][ci] += 1.0
|
||||
g.bandN[band] += 1.0
|
||||
inc g.trained
|
||||
|
||||
proc lgApplyDecay(g: var LeadGainGun) =
|
||||
## Forgetting for `TR_LEADGAIN_MEM=decay`: shrink the hit counts and, more
|
||||
## strongly, pull them toward the gain-1.0 column so stale evidence ages out.
|
||||
let f = 1.0 - g.decayFrac
|
||||
if f >= 1.0: return
|
||||
for b in 0 ..< LG_NBANDS:
|
||||
for ci in 0 ..< g.cands.len:
|
||||
g.bandHits[b][ci] *= f
|
||||
g.bandN[b] *= f
|
||||
inc g.decays
|
||||
|
||||
proc lgGainFor(g: LeadGainGun, band: int): float =
|
||||
## The band's gain is the candidate with the highest observed hit rate.
|
||||
## Ties keep the SMALLER candidate (the scan is ascending), which is the
|
||||
## conservative choice for the long-range regime this corrector targets.
|
||||
## Returns 1.0 (Pattern) below the range gate or when the band is cold.
|
||||
if band < LG_GAIN_BAND_MIN: return 1.0
|
||||
if g.cands.len == 0: return 1.0
|
||||
# A single candidate is a FIXED gain: apply it from the first shot, never
|
||||
# consult the counts. This is the no-learning arm of the live sweep.
|
||||
if g.cands.len == 1: return g.cands[0]
|
||||
if g.bandN[band] < float(g.minObs): return 1.0
|
||||
var best = -1
|
||||
var bestRate = -1.0
|
||||
for ci in 0 ..< g.cands.len:
|
||||
let rate = g.bandHits[band][ci] / g.bandN[band]
|
||||
if rate > bestRate:
|
||||
bestRate = rate
|
||||
best = ci
|
||||
if best < 0: return 1.0
|
||||
g.cands[best]
|
||||
|
||||
# ── deferred-label resolution (prequential learning) ─────────────────────────
|
||||
|
||||
proc resolvePending(g: var LeadGainGun, state: WorldState) =
|
||||
var w = 0
|
||||
for i in 0 ..< g.pendingCount:
|
||||
let p = g.pending[i]
|
||||
let due = p.fireTick + p.horizon
|
||||
if due > state.tick:
|
||||
g.pending[w] = p
|
||||
inc w
|
||||
elif due == state.tick:
|
||||
let obs = tmhObservedAt(g.tmh, state.tick, p.selfX, p.selfY)
|
||||
if obs.ok and (state.tick - obs.lastSeenTick) <= TMH_STALE_MAX:
|
||||
let err = wrapRadLg(obs.bearing - p.baseBearing)
|
||||
let reqLead = wrapRadLg(err + p.lead)
|
||||
g.lgAccumulate(radToDeg(p.lead), radToDeg(reqLead), p.tolDeg, p.band)
|
||||
inc g.sinceDecay
|
||||
if g.memMode == lgDecay and g.sinceDecay >= g.decayEvery:
|
||||
g.lgApplyDecay()
|
||||
g.sinceDecay = 0
|
||||
else:
|
||||
inc g.pendingDropped
|
||||
else:
|
||||
inc g.pendingDropped
|
||||
g.pendingCount = w
|
||||
|
||||
# ── logging ──────────────────────────────────────────────────────────────────
|
||||
|
||||
proc lgLog(g: var LeadGainGun, state: WorldState, band: int, gain, leadDeg: float) =
|
||||
## ONE change-gated `[lg]` line (behind TR_LEADGAIN_LOG=1) so a user tailing
|
||||
## the GUI log sees the gain the corrector is applying. The APPLIED gain and
|
||||
## the resulting angular `shift` are both on the line: the boot report proves
|
||||
## the knob reached the process, this proves the gun actually used it.
|
||||
if not g.logEnabled: return
|
||||
let shiftDeg = (gain - 1.0) * leadDeg
|
||||
let key = fmt"{gain:.2f}|{band}"
|
||||
if key == g.lastLogKey: return
|
||||
g.lastLogKey = key
|
||||
var rate = 0.0
|
||||
for ci in 0 ..< g.cands.len:
|
||||
if abs(g.cands[ci] - gain) < 1e-9: rate = g.bandHits[band][ci] / max(1.0, g.bandN[band])
|
||||
echo fmt"[lg] t={state.tick} band={LG_BAND_LO[band]:.0f}+ gain={gain:.2f} " &
|
||||
fmt"shift={shiftDeg:+.2f}deg rate={rate:.3f} n={g.bandN[band]:.0f} " &
|
||||
fmt"ncand={g.cands.len} trained={g.trained} " &
|
||||
fmt"pend={g.pendingCount} dropped={g.pendingDropped} mode={memModeName(g.memMode)}"
|
||||
|
||||
# ── reset hooks (mirroring TmHorizonGun) ─────────────────────────────────────
|
||||
|
||||
proc resetRound(g: var LeadGainGun) =
|
||||
## PER-ROUND reset: observation ring, deferred labels and per-tick caches (the
|
||||
## bots teleport between rounds). The gain counts are deliberately KEPT — they
|
||||
## are battle-scale and a new round is not a new enemy.
|
||||
g.tmh.resetRoundState()
|
||||
g.pendingCount = 0
|
||||
g.lastTick = -1
|
||||
g.lastEnqTick = -1
|
||||
g.lastEnqBucket = -1
|
||||
g.lastLogKey = ""
|
||||
|
||||
proc resetRoundState*(g: var LeadGainGun) =
|
||||
if not g.initialized: return
|
||||
g.resetRound()
|
||||
|
||||
proc resetLearning*(g: var LeadGainGun, reason = "") =
|
||||
## PER-BATTLE / PER-ENEMY wipe: gain counts, counters and the round state.
|
||||
if not g.initialized: return
|
||||
for b in 0 ..< LG_NBANDS:
|
||||
for ci in 0 ..< g.cands.len: g.bandHits[b][ci] = 0.0
|
||||
g.bandN[b] = 0.0
|
||||
g.lastGain[b] = 1.0
|
||||
g.trained = 0
|
||||
g.sinceDecay = 0
|
||||
g.decays = 0
|
||||
g.corrections = 0
|
||||
g.observedTargetId = -1
|
||||
g.resetRound()
|
||||
if reason.len > 0 and g.logEnabled:
|
||||
echo fmt"[lg-reset] reason={reason}"
|
||||
|
||||
proc targetChanged*(g: var LeadGainGun, enemyId: int): bool =
|
||||
## Per-ENEMY reset: wipe when the target changes to a different bot id. First
|
||||
## acquisition never wipes, so the round-start pick does not cold-start us.
|
||||
if not g.resetOnTarget: return false
|
||||
if enemyId < 0: return false
|
||||
if g.observedTargetId >= 0 and enemyId != g.observedTargetId:
|
||||
g.resetLearning("target_change")
|
||||
g.observedTargetId = enemyId
|
||||
return true
|
||||
g.observedTargetId = enemyId
|
||||
false
|
||||
|
||||
# ── Gun interface ────────────────────────────────────────────────────────────
|
||||
|
||||
proc isWarmedUp*(g: LeadGainGun): bool {.inline.} = true
|
||||
|
||||
proc networkBytes*(g: LeadGainGun): int =
|
||||
## No neural network is held any more; kept for the boot report / guard test.
|
||||
0
|
||||
|
||||
proc predict*(g: var LeadGainGun, state: WorldState,
|
||||
bulletSpeed: float): GunPrediction =
|
||||
g.ensureInit()
|
||||
|
||||
# Round boundary: a tick regression means a new round.
|
||||
if state.tick < g.lastTick: g.resetRound()
|
||||
|
||||
# Once per tick: observe the world, then resolve any labels now due.
|
||||
if state.tick != g.lastTick:
|
||||
tmhUpdateHistory(g.tmh, state)
|
||||
g.resolvePending(state)
|
||||
g.lastTick = state.tick
|
||||
|
||||
# The base prediction is Pattern; LEADGAIN only scales its lead over LOS.
|
||||
let base = g.tmh.pattern.predict(state, bulletSpeed)
|
||||
if bulletSpeed <= 0.0: return base
|
||||
|
||||
let dist = hypot(state.enemyX - state.selfX, state.enemyY - state.selfY)
|
||||
let h = tmhHorizonFor(dist, bulletSpeed)
|
||||
let hb = tmhHorizonBucket(h)
|
||||
let band = lgBandOf(dist)
|
||||
|
||||
let los = arctan2(state.enemyY - state.selfY, state.enemyX - state.selfX)
|
||||
let baseBearing = arctan2(base.y - state.selfY, base.x - state.selfX)
|
||||
let lead = wrapRadLg(baseBearing - los)
|
||||
|
||||
# Enqueue one deferred sample per (tick, horizon bucket): `predict` runs once
|
||||
# per power bin, so all four horizons contribute evidence.
|
||||
if g.lastEnqTick != state.tick or g.lastEnqBucket != hb:
|
||||
if g.pendingCount < LG_PENDING_CAP:
|
||||
g.pending[g.pendingCount] = LgPending(
|
||||
fireTick: state.tick, horizon: h, band: band,
|
||||
selfX: state.selfX, selfY: state.selfY,
|
||||
baseBearing: baseBearing, lead: lead, tolDeg: lgTolDeg(dist))
|
||||
inc g.pendingCount
|
||||
else:
|
||||
inc g.pendingDropped
|
||||
g.lastEnqTick = state.tick
|
||||
g.lastEnqBucket = hb
|
||||
|
||||
# Readout: a fractional gain may be BELOW 1.0. When cold / gated out the
|
||||
# learner returns 1.0 and the base prediction is returned unchanged.
|
||||
let gain = g.lgGainFor(band)
|
||||
g.lastGain[band] = gain
|
||||
if abs(gain - 1.0) < 1e-9: return base
|
||||
inc g.corrections
|
||||
g.lgLog(state, band, gain, radToDeg(lead))
|
||||
tmhApplyShift(state.selfX, state.selfY, base.x, base.y,
|
||||
radToDeg((gain - 1.0) * lead))
|
||||
|
||||
proc onResult*(g: var LeadGainGun, e: FeedbackEvent) =
|
||||
## Labels come from our own observation ring, not from virtual-bullet
|
||||
## feedback, so there is nothing to do here. The hook exists for the rack.
|
||||
discard
|
||||
Reference in New Issue
Block a user