T6: Extract Tsetlin Machine as gun module #184

New Issue

2026-09-19T14:49:12+02:00

SirStone commented

2026-09-19 14:49:12 +02:00

Parent

#165 (ModularBot wayfinder map)

What to build

Extract TM predictor from BNNBot_garage and adapt it as a Gun module. The TM gun is a learning gun — it uses onResult() to train on miss corrections.

Extract tsetlin_predictor.nim and binary_encoding.nim into a gun module at common_libs/guns/tsetlin.nim
Adapt to Gun interface: predict() runs TM forward pass + correction on extrapolated position, onResult() matches tick via internal eligibility traces and calls learn()
Manages own frame history window
Configurable: clauses, T, s, states

Register in ModularBot. The learning gun competes with analytical ones — should eventually outperform against complex movement patterns.

Acceptance criteria

TM gun implements Gun interface
predict() returns absolute (x, y) using TM correction on linear extrapolation
onResult() triggers TM learning via eligibility traces
Registered in ModularBot — four guns competing
TM gun's fitness improves over rounds (visible in virtual bullet stats)

Blocked by

#179 (T1)

## Parent #165 (ModularBot wayfinder map) ## What to build Extract TM predictor from BNNBot_garage and adapt it as a Gun module. The TM gun is a learning gun — it uses onResult() to train on miss corrections. - Extract `tsetlin_predictor.nim` and `binary_encoding.nim` into a gun module at `common_libs/guns/tsetlin.nim` - Adapt to Gun interface: predict() runs TM forward pass + correction on extrapolated position, onResult() matches tick via internal eligibility traces and calls learn() - Manages own frame history window - Configurable: clauses, T, s, states Register in ModularBot. The learning gun competes with analytical ones — should eventually outperform against complex movement patterns. ## Acceptance criteria - [ ] TM gun implements Gun interface - [ ] predict() returns absolute (x, y) using TM correction on linear extrapolation - [ ] onResult() triggers TM learning via eligibility traces - [ ] Registered in ModularBot — four guns competing - [ ] TM gun's fitness improves over rounds (visible in virtual bullet stats) ## Blocked by #179 (T1)

SirStone added the ready-for-agent label 2026-09-19 14:49:12 +02:00

SirStone referenced this issue

2026-09-19 14:49:33 +02:00

T7: Battle test ModularBot against varied enemies #185

SirStone referenced this issue

2026-09-19 14:50:13 +02:00

ModularBot — Modular multi-gun bot with pluggable harnesses #165

SirStone commented

2026-09-19 15:45:55 +02:00

TsetlinGun implemented and wired into ModularBot. Build passes.

What was done:

Created common_libs/guns/tsetlin.nim — self-contained: binary encoding and TM predictor inlined, no dependency on BNNBot_garage.
Implements Gun interface: predict(state, bulletSpeed) → GunPrediction and onResult(FeedbackEvent).
predict(): encodes WorldState into 870-bit binary vector (sliding window of 10 frames × 83 bits + 40 self bits), runs TM forward pass, applies (cx, cy) correction on top of linear extrapolation, returns clamped absolute (x, y). Falls back to pure linear during warm-up (first 10 frames).
onResult(): matches the resolved bullet back to its stored trace via prediction coords, computes shaped residual from miss distance, calls TM learn on both output dimensions.
Internal ring buffer of 64 trace slots for eligibility matching.
TM config: 50 clauses, T=25.0, s=1.5, states=32 (int16).
Registered in ModularBot as gun 2 (tracker: vb.initTracker(3)); headOn=0, linear=1, tsetlin=2.

Design notes:

onResult receives miss distance but not direction. Residual is split equally across x/y axes (√2 factor). Upgrade path: add enemyX/Y to FeedbackEvent for directional residuals.
circular import stays in ModularBot (other work), CircularGun field retained but not wired to tracker yet.
ponytail: comment in tsetlin.nim marks the 50/50 residual split as a deliberate simplification.

TsetlinGun implemented and wired into ModularBot. Build passes. **What was done:** - Created `common_libs/guns/tsetlin.nim` — self-contained: binary encoding and TM predictor inlined, no dependency on BNNBot_garage. - Implements Gun interface: `predict(state, bulletSpeed) → GunPrediction` and `onResult(FeedbackEvent)`. - `predict()`: encodes WorldState into 870-bit binary vector (sliding window of 10 frames × 83 bits + 40 self bits), runs TM forward pass, applies (cx, cy) correction on top of linear extrapolation, returns clamped absolute (x, y). Falls back to pure linear during warm-up (first 10 frames). - `onResult()`: matches the resolved bullet back to its stored trace via prediction coords, computes shaped residual from miss distance, calls TM learn on both output dimensions. - Internal ring buffer of 64 trace slots for eligibility matching. - TM config: 50 clauses, T=25.0, s=1.5, states=32 (int16). - Registered in ModularBot as gun 2 (`tracker: vb.initTracker(3)`); headOn=0, linear=1, tsetlin=2. **Design notes:** - `onResult` receives miss distance but not direction. Residual is split equally across x/y axes (√2 factor). Upgrade path: add `enemyX/Y` to `FeedbackEvent` for directional residuals. - `circular` import stays in ModularBot (other work), `CircularGun` field retained but not wired to tracker yet. - `ponytail:` comment in tsetlin.nim marks the 50/50 residual split as a deliberate simplification.

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: SirStone/SirRoboGarage#184