SirStone/SirRoboGarage
1
0
Fork 0
Code Issues 28 Pull Requests Actions Projects Releases Wiki Activity

fix(SNNBot): average matured snapshot targets instead of multiple updates

Browse Source 
Multiple superSpikeUpdate calls per EVALUATE caused oscillation by
effectively multiplying the learning rate. Now: accumulate retroTarget
via circular mean across all matured snapshots, apply one update with
the averaged target and most recent snapshot's SNN state.
This commit is contained in:
Davide Cappellini
2026-09-14 22:28:41 +02:00
parent d5777fba07
commit 42b6cebd8e
1 changed files with 21 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
SNNBot_garage/src/SNNBot.nim
+21 -4
View File
@@ -464,9 +464,16 @@ method run*(bot: SNNBot) =
# Retroactive would-have-hit error signal: find the most recently matured snapshot. # Retroactive would-have-hit error signal: find the most recently matured snapshot.
# A snapshot matures when currentTick >= snapshotTick + ceil(distance / BULLET_SPEED). # A snapshot matures when currentTick >= snapshotTick + ceil(distance / BULLET_SPEED).
# Look up enemy position at impact tick from ring buffer (offset from most-recent write). # Look up enemy position at impact tick from ring buffer (offset from most-recent write).
var retroTarget = bot.lastRelBearing # fallback; overwritten by each matured snapshot (last one used for logging) var retroTarget = bot.lastRelBearing # fallback; overwritten after averaging (used for logging)
var hasMatured = false var hasMatured = false
var keepIdx = 0 # first non-matured snapshot to keep var keepIdx = 0 # first non-matured snapshot to keep
# Accumulate retroTarget circular components across all matured snapshots.
var sumSin = 0.0; var sumCos = 0.0
var matureCount = 0
# Most-recent matured snapshot's SNN state (most relevant for weight update).
var bestSpikes: array[N_HID, float]
var bestVSnap: array[N_HID, float]
var bestPreTrace: array[N_IN, float]
for i in 0 ..< bot.snapshots.len: for i in 0 ..< bot.snapshots.len:
let snap = bot.snapshots[i] let snap = bot.snapshots[i]
let travelTicks = int(ceil(snap.distance / BULLET_SPEED)) let travelTicks = int(ceil(snap.distance / BULLET_SPEED))
@@ -483,13 +490,23 @@ method run*(bot: SNNBot) =
let ex = bot.posBuf[foundSlot].x let ex = bot.posBuf[foundSlot].x
let ey = bot.posBuf[foundSlot].y let ey = bot.posBuf[foundSlot].y
let absBearing = directionTo(snap.botX, snap.botY, ex, ey) let absBearing = directionTo(snap.botX, snap.botY, ex, ey)
retroTarget = normalizeRelativeAngle(absBearing - snap.gunHeading) let rt = normalizeRelativeAngle(absBearing - snap.gunHeading)
# Update weights for every matured snapshot immediately; earlier ones were previously discarded. # Accumulate circular mean components
bot.snn.superSpikeUpdate(snap.spikes, snap.vSnap, snap.preTrace, retroTarget) sumSin += sin(degToRad(rt))
sumCos += cos(degToRad(rt))
inc matureCount
# Keep most-recent (highest index) snapshot's SNN state
bestSpikes = snap.spikes
bestVSnap = snap.vSnap
bestPreTrace = snap.preTrace
hasMatured = true hasMatured = true
keepIdx = i + 1 # discard matured snapshots up to and including this one keepIdx = i + 1 # discard matured snapshots up to and including this one
else: else:
break # snapshots are in order; stop at first non-matured break # snapshots are in order; stop at first non-matured
# One update per EVALUATE using circular-mean target — stable learning rate regardless of snapshot count.
if matureCount > 0:
retroTarget = arctan2(sumSin, sumCos) * 180.0 / PI
bot.snn.superSpikeUpdate(bestSpikes, bestVSnap, bestPreTrace, retroTarget)
# Discard matured snapshots; immature ones are preserved automatically (keepIdx stays 0 # Discard matured snapshots; immature ones are preserved automatically (keepIdx stays 0
# or points past the last matured entry; the rest of bot.snapshots is kept intact). # or points past the last matured entry; the rest of bot.snapshots is kept intact).
if keepIdx > 0: if keepIdx > 0: