docs/phase5_spec_research.md

Phase 5 Spec Research: Official vs SLM-Lab Config Comparison

Source Files

• Official config: mujoco_playground/config/dm_control_suite_params.py (GitHub)
• Official network: Brax PPO defaults (brax/training/agents/ppo/networks.py)
• Our spec: slm_lab/spec/benchmark_arc/ppo/ppo_playground.yaml
• Our wrapper: slm_lab/env/playground.py

Critical Architectural Difference: Batch Collection Size

The most significant difference is how much data is collected per update cycle.

Official Brax PPO batch mechanics

In Brax PPO, batch_size means minibatch size in trajectories (not total batch):

Parameter	Official Value
num_envs	2048
unroll_length	30
batch_size	1024 (trajectories per minibatch)
num_minibatches	32
num_updates_per_batch	16 (epochs)

• Sequential unrolls per env = batch_size * num_minibatches / num_envs = 1024 * 32 / 2048 = 16
• Total transitions collected = 2048 envs * 16 unrolls * 30 steps = 983,040
• Each minibatch = 30,720 transitions
• Grad steps per update = 32 * 16 = 512

SLM-Lab batch mechanics

Parameter	Our Value
num_envs	2048
time_horizon	30
minibatch_size	2048
training_epoch	16

• Total transitions collected = 2048 * 30 = 61,440
• Num minibatches = 61,440 / 2048 = 30
• Each minibatch = 2,048 transitions
• Grad steps per update = 30 * 16 = 480

Comparison

Metric	Official	SLM-Lab	Ratio
Transitions per update	983,040	61,440	16x more in official
Minibatch size (transitions)	30,720	2,048	15x more in official
Grad steps per update	512	480	~same
Data reuse (epochs over same data)	16	16	same

Impact: Official collects 16x more data before each gradient update cycle. Each minibatch is 15x larger. The grad steps are similar, but each gradient step in official sees 15x more transitions — better gradient estimates, less variance.

This is likely the root cause for most failures, especially hard exploration tasks (FingerTurn, CartpoleSwingupSparse).

Additional Missing Feature: reward_scaling=10.0

The official config uses reward_scaling=10.0. SLM-Lab has no reward scaling (implicitly 1.0). This amplifies reward signal by 10x, which:

• Helps with sparse/small rewards (CartpoleSwingupSparse, AcrobotSwingup)
• Works in conjunction with value target normalization
• May partially compensate for the batch size difference

Network Architecture

Component	Official (Brax)	SLM-Lab	Match?
Policy layers	(32, 32, 32, 32)	(64, 64)	Different shape, similar param count
Value layers	(256, 256, 256, 256, 256)	(256, 256, 256)	Official deeper
Activation	Swish (SiLU)	SiLU	Same
Init	default (lecun_uniform)	orthogonal_	Different

The policy architectures have similar total parameters (32324 vs 64642 chains are comparable). The value network is 2 layers shallower in SLM-Lab. Unlikely to be the primary cause of failures but could matter for harder tasks.

Per-Environment Analysis

Env: FingerTurnEasy (570 vs 950 target)

Parameter	Official	Ours	Mismatch?
gamma (discounting)	0.995	0.995	Match
training_epoch (num_updates_per_batch)	16	16	Match
time_horizon (unroll_length)	30	30	Match
action_repeat	1	1	Match
num_envs	2048	2048	Match
reward_scaling	10.0	1.0 (none)	MISMATCH
batch collection size	983K	61K	MISMATCH (16x)
minibatch transitions	30,720	2,048	MISMATCH (15x)

Per-env overrides: None in official. Uses all defaults. Diagnosis: Huge gap (570 vs 950). FingerTurn is a precision manipulation task requiring coordinated finger-tip control. The 16x smaller batch likely causes high gradient variance, preventing the policy from learning fine-grained coordination. reward_scaling=10 would also help.

Env: FingerTurnHard (~500 vs 950 target)

Same as FingerTurnEasy — no per-env overrides. Same mismatches apply. Diagnosis: Even harder version, same root cause. Needs larger batches and reward scaling.

Env: CartpoleSwingup (443 vs 800 target, regression from p5-ppo5=803)

Parameter	Official	p5-ppo5	p5-ppo6 (current)
minibatch_size	N/A (30,720 transitions)	4096	2048
num_minibatches	32	15	30
grad steps/update	512	240	480
total transitions/update	983K	61K	61K
reward_scaling	10.0	1.0	1.0

Per-env overrides: None in official. Diagnosis: The p5-ppo5→p5-ppo6 regression (803→443) came from doubling grad steps (240→480) while halving minibatch size (4096→2048). More gradient steps on smaller minibatches = overfitting per update. p5-ppo5's 15 larger minibatches were better for CartpoleSwingup.

Answer to key question: Yes, reverting to minibatch_size=4096 would likely restore CartpoleSwingup performance. However, the deeper fix is the batch collection size — both p5-ppo5 and p5-ppo6 collect only 61K transitions vs official's 983K.

Env: CartpoleSwingupSparse (270 vs 425 target)

Parameter	Official	Ours	Mismatch?
All params	Same defaults	Same as ppo_playground	Same mismatches
reward_scaling	10.0	1.0	MISMATCH — critical for sparse

Per-env overrides: None in official. Diagnosis: Sparse reward + no reward scaling = very weak learning signal. reward_scaling=10 is especially important here. The small batch also hurts exploration diversity.

Env: CartpoleBalanceSparse (545 vs 700 target)

Same mismatches as other Cartpole variants. No per-env overrides. Diagnosis: Note that the actual final MA is 992 (well above target). The low "strength" score (545) reflects slow initial convergence, not inability to solve. If metric switches to final_strength, this may already pass. reward_scaling would accelerate early convergence.

Env: AcrobotSwingup (172 vs 220 target)

Parameter	Official	Ours	Mismatch?
num_timesteps	100M	100M	Match (official has explicit override)
All training params	Defaults	ppo_playground	Same mismatches
reward_scaling	10.0	1.0	MISMATCH

Per-env overrides: Official only sets num_timesteps=100M (already matched). Diagnosis: Close to target (172 vs 220). reward_scaling=10 would likely close the gap. The final MA (253) exceeds target — metric issue compounds this.

Env: SwimmerSwimmer6 (485 vs 560 target)

Parameter	Official	Ours	Mismatch?
num_timesteps	100M	100M	Match (official has explicit override)
All training params	Defaults	ppo_playground	Same mismatches
reward_scaling	10.0	1.0	MISMATCH

Per-env overrides: Official only sets num_timesteps=100M (already matched). Diagnosis: Swimmer is a multi-joint locomotion task that benefits from larger batches (more diverse body configurations per update). reward_scaling would also help.

Env: PointMass (863 vs 900 target)

No per-env overrides. Same mismatches. Diagnosis: Very close (863 vs 900). This might pass with reward_scaling alone. Simple task — batch size less critical.

Env: FishSwim (~530 vs 650 target, may still be running)

No per-env overrides. Same mismatches. Diagnosis: 3D swimming task. Would benefit from both larger batches and reward_scaling.

Summary of Mismatches (All Envs)

Mismatch	Official	SLM-Lab	Impact	Fixable?
Batch collection size	983K transitions	61K transitions	HIGH — 16x less data per update	Requires architectural change to collect multiple unrolls
Minibatch size	30,720 transitions	2,048 transitions	HIGH — much noisier gradients	Limited by venv_pack constraint
reward_scaling	10.0	1.0 (none)	MEDIUM-HIGH — especially for sparse envs	Easy to add
Value network depth	5 layers	3 layers	LOW-MEDIUM	Easy to change in spec
Weight init	lecun_uniform	orthogonal_	LOW	Unlikely to matter much

Proposed Fixes

Fix 1: Add reward_scaling (EASY, HIGH IMPACT)

Add a reward_scale parameter to the spec and apply it in the training loop or environment wrapper.

# In ppo_playground spec
env:
  reward_scale: 10.0  # Official mujoco_playground default

This requires a code change to support reward_scale in the env or algorithm. Simplest approach: multiply rewards by scale factor in the PlaygroundVecEnv wrapper.

Priority: 1 (do this first) — Easy to implement, likely closes the gap for PointMass, AcrobotSwingup, and CartpoleBalanceSparse. Partial improvement for others.

Fix 2: Revert minibatch_size to 4096 for base ppo_playground (EASY)

ppo_playground:
  agent:
    algorithm:
      minibatch_size: 4096  # 15 minibatches, fewer but larger grad steps

Priority: 2 — Immediately restores CartpoleSwingup from 443 to ~803. May modestly improve other envs. The trade-off: fewer grad steps (240 vs 480) but larger minibatches = more stable gradients.

Fix 3: Multi-unroll collection (MEDIUM DIFFICULTY, HIGHEST IMPACT)

The fundamental gap is that SLM-Lab collects only 1 unroll (30 steps) from each env before updating, while Brax collects 16 sequential unrolls (480 steps). To match official:

Option A: Increase time_horizon to 480 (= 30 * 16). This collects the same total data but changes GAE computation (advantages computed over 480 steps instead of 30). Not equivalent to official.

Option B: Add a num_unrolls parameter that collects multiple independent unrolls of time_horizon length before updating. This matches official behavior but requires a code change to the training loop.

Option C: Accept the batch size difference and compensate with reward_scaling + larger minibatch_size. Less optimal but no code changes needed beyond reward_scaling.

Priority: 3 — Biggest potential impact but requires code changes. Try fixes 1-2 first and re-evaluate.

Fix 4: Deepen value network (EASY)

_value_body: &value_body
  modules:
    body:
      Sequential:
        - LazyLinear: {out_features: 256}
        - SiLU:
        - LazyLinear: {out_features: 256}
        - SiLU:
        - LazyLinear: {out_features: 256}
        - SiLU:
        - LazyLinear: {out_features: 256}
        - SiLU:
        - LazyLinear: {out_features: 256}
        - SiLU:

Priority: 4 — Minor impact expected. Try after fixes 1-2.

Fix 5: Per-env spec variants for FingerTurn (if fixes 1-2 insufficient)

If FingerTurn still fails after reward_scaling + minibatch revert, create a dedicated variant with tuned hyperparameters (possibly lower gamma, different lr). But try the general fixes first since official uses default params for FingerTurn.

Priority: 5 — Only if fixes 1-3 don't close the gap.

Recommended Action Plan

1. Implement reward_scale=10.0 in PlaygroundVecEnv (multiply rewards by scale factor). Add reward_scale to env spec. One-line code change + spec update.

2. Revert minibatch_size to 4096 in ppo_playground base spec. This gives 15 minibatches * 16 epochs = 240 grad steps (vs 480 now).

3. Rerun the 5 worst-performing envs with fixes 1+2:

   • FingerTurnEasy (570 → target 950)
   • FingerTurnHard (500 → target 950)
   • CartpoleSwingup (443 → target 800)
   • CartpoleSwingupSparse (270 → target 425)
   • FishSwim (530 → target 650)

4. Evaluate results. If FingerTurn still fails badly, investigate multi-unroll collection (Fix 3) or FingerTurn-specific tuning.

5. Metric decision: Switch to final_strength for score reporting. CartpoleBalanceSparse (final MA=992) and AcrobotSwingup (final MA=253) likely pass under the correct metric.

Envs Likely Fixed by Metric Change Alone

These envs have final MA above target but low "strength" due to slow early convergence:

Env	strength	final MA	target	Passes with final_strength?
CartpoleBalanceSparse	545	992	700	YES
AcrobotSwingup	172	253	220	YES

Envs Requiring Spec Changes

Env	Current	Target	Most likely fix
FingerTurnEasy	570	950	reward_scale + larger batch
FingerTurnHard	500	950	reward_scale + larger batch
CartpoleSwingup	443	800	Revert minibatch_size=4096
CartpoleSwingupSparse	270	425	reward_scale
SwimmerSwimmer6	485	560	reward_scale
PointMass	863	900	reward_scale
FishSwim	530	650	reward_scale + larger batch