Network and state vector dimension changes #26

New Issue

2026-08-17T18:55:27+02:00

SirStone commented

2026-08-17 18:55:27 +02:00

Parent

#24 — PRD: Command abstraction layer for PPO action space

What to build

Mechanical dimension changes across the network and state vector modules to support the new 6-dim action space and 44-dim state input.

Network changes (in network.nim):

Actor output layer: 5 → 6
Learnable logStd: 5 → 6
Input layer: 42 → 44
Discard all saved weights (shapes change) — delete or move existing weight files

State vector changes (in state_vector.nim):

Append 2 new inputs at the end of the state vector:
- Remaining distance to goto target, normalized by battlefield diagonal → [0, 1]
- Remaining gun angle to aimTo target, normalized by 180° → [0, 1]
buildStateVector needs to accept these two values as parameters (they come from PPO_Bot at runtime)
Update dimension constant from 42 to 44

Training changes (in training.nim):

Update any hardcoded 5 references in log-prob computation and dLogP_dMean allocation to 6

Acceptance criteria

actorForward produces 6-dim output tensor
criticForward accepts 44-dim input tensor
logStd is 6-dimensional
buildStateVector returns 44-dim tensor
State dims 42 and 43 contain normalized remaining goto distance and gun angle
Old weight files are removed or archived
computeLogProb works with 6-dim actions
Bot compiles and network initializes without errors

Blocked by

None — can start immediately

## Parent #24 — PRD: Command abstraction layer for PPO action space ## What to build Mechanical dimension changes across the network and state vector modules to support the new 6-dim action space and 44-dim state input. **Network changes** (in `network.nim`): - Actor output layer: 5 → 6 - Learnable `logStd`: 5 → 6 - Input layer: 42 → 44 - Discard all saved weights (shapes change) — delete or move existing weight files **State vector changes** (in `state_vector.nim`): - Append 2 new inputs at the end of the state vector: - Remaining distance to goto target, normalized by battlefield diagonal → [0, 1] - Remaining gun angle to aimTo target, normalized by 180° → [0, 1] - `buildStateVector` needs to accept these two values as parameters (they come from PPO_Bot at runtime) - Update dimension constant from 42 to 44 **Training changes** (in `training.nim`): - Update any hardcoded `5` references in log-prob computation and `dLogP_dMean` allocation to `6` ## Acceptance criteria - [ ] `actorForward` produces 6-dim output tensor - [ ] `criticForward` accepts 44-dim input tensor - [ ] `logStd` is 6-dimensional - [ ] `buildStateVector` returns 44-dim tensor - [ ] State dims 42 and 43 contain normalized remaining goto distance and gun angle - [ ] Old weight files are removed or archived - [ ] `computeLogProb` works with 6-dim actions - [ ] Bot compiles and network initializes without errors ## Blocked by None — can start immediately

SirStone added the ready-for-agent label 2026-08-17 18:55:27 +02:00

SirStone referenced this issue

2026-08-17 18:55:41 +02:00

Action decoding rewrite for command abstraction #27

SirStone referenced this issue

2026-08-17 18:55:52 +02:00

PPO_Bot wiring and integration test #28

SirStone commented

2026-08-17 19:09:10 +02:00