hyperparameter-optimization

name	hyperparameter-optimization
description	Unified PPO hyperparameter and reward/penalty weight search for VBot navigation. Grid, random, and Bayesian optimization across learning rate, network architecture, training dynamics, and reward scales.

Purpose

Unified hyperparameter tuning for PPO training — reward/penalty weights ARE hyperparameters and are searched in the same loop:

• PPO dynamics (learning rate, entropy, clipping, epochs)
• Network architecture (layer sizes)
• Reward/penalty weights (navigation rewards, stability penalties, termination)
• Grid/random/Bayesian search strategies
• Parameter importance analysis
• Best config export and comparison

🔴 AutoML-First Policy (MANDATORY): NEVER use train.py for parameter search. The AutoML pipeline (automl.py) handles unified HP+reward search. See .github/copilot-instructions.md for the full policy.

train.py is ONLY for: smoke tests (<500K steps), --render visual debug, or final deployment runs.

Core Design Principle: Reward weights and PPO parameters interact. Searching them separately is suboptimal — a high termination penalty needs a different learning rate than a low one. The AutoML pipeline searches both jointly in a single trial loop.

IMPORTANT — Operational Guardrails:

• The AutoML pipeline (automl.py) handles unified HP+reward search automatically. Do NOT re-implement search logic.
• sample_from_space() already returns native Python types (int, float). The numpy serialization bug is fixed.
• To run search, use the command in Quick Start below. Do not re-read or re-inspect automl.py.

Related Skills:

• training-pipeline — Hub with Quick Start (start here)
• reward-penalty-engineering — Exploration methodology for what to search (diagnosis → hypothesis → test cycle)
• training-campaign — Execute search experiments
• curriculum-learning — Stage-specific tuning

Quick Start

# Launch AutoML unified search (HP + reward weights)
uv run starter_kit_schedule/scripts/automl.py `
    --mode stage `
    --budget-hours 12 `
    --hp-trials 8

# Monitor search progress
Get-Content starter_kit_schedule/progress/automl_state.yaml

When to Use

Task	Use This
Tune learning rate	✅
Search network architectures	✅
Optimize PPO dynamics	✅
Compare search methods	✅
Search reward/penalty weights	✅ (built-in — every trial samples reward scales too)
Diagnose bad behavior to decide what to add/remove	❌ → Use reward-penalty-engineering

Unified Search Space

Each trial samples both PPO parameters and reward weights. This is the complete search space:

PPO Parameters

Parameter	Type	Range	Default	Impact
learning_rate	log-uniform	1e-5 – 1e-3	3e-4	Critical — convergence speed
discount_factor	choice	0.95–0.999	0.99	Credit assignment horizon
lambda_param	choice	0.9–0.99	0.95	GAE bias-variance
ratio_clip	choice	0.1–0.3	0.2	Policy update magnitude
entropy_loss_scale	log-uniform	1e-4 – 1e-2	1e-3	Exploration strength
learning_epochs	choice	4, 5, 6, 8	5	Updates per rollout
mini_batches	choice	16, 32, 64	32	Batch size
rollouts	choice	16, 24, 32, 48	24	Steps before update

Network Architecture Options

Config	Layers	Parameters	Use Case
Small	[128, 64]	~20K	Fast iteration
Medium	[256, 128, 64]	~80K	Default
Large	[512, 256, 128]	~200K	Complex policies
Deep	[256, 256, 256]	~150K	Feature extraction

Reward / Penalty Weights

Reward weights come from starter_kit/{task}/vbot/cfg.py → RewardConfig.scales. Each trial samples a complete reward weight vector jointly with PPO params.

Concrete reward weight ranges, default values, and known warnings (sprint-crash, deceleration moat, lazy robot) are task-specific. See: starter_kit_docs/{task-name}/Task_Reference.md → "Reward Search Space" section.

General schema for reward weight search entries:

reward_weights:
  <reward_name>:
    type: "uniform" | "loguniform" | "choice"
    low: <min_value>     # for uniform/loguniform
    high: <max_value>    # for uniform/loguniform
    values: [...]        # for choice
    prior: <default>     # center of search
    category: "navigation" | "stability" | "efficiency" | "terminal"

Key principle: arrival_bonus (or equivalent goal reward) must be large enough to dominate per-step bonuses like alive_bonus × typical_episode_length. See .github/copilot-instructions.md → "Reward Budget Audit Principle".

Search Space Schema

# starter_kit_schedule/templates/ppo_search_space.yaml
search_space_id: "unified_navigation_v1"

# Fixed parameters
fixed:
  seed: 42
  num_envs: 2048
  grad_norm_clip: 1.0
  ratio_clip: 0.2
  value_clip: 0.2

# Searchable PPO parameters
ppo_params:
  learning_rate:
    type: "loguniform"
    low: 1e-5
    high: 1e-3
    prior: 3e-4
    
  policy_hidden_layer_sizes:
    type: "categorical"
    choices:
      - [128, 64]
      - [256, 128, 64]
      - [512, 256, 128]
      - [256, 256, 256]
    prior: [256, 128, 64]
    
  rollouts:
    type: "choice"
    values: [16, 24, 32, ]
     
    
  
     
     [, , , ]
     
    
  
     
     
     
     



  
     
     
     
     
    
  
     
     
     
     
    
  
     
     
     
     
    
  
     
     
     
     
    
  
     
     
     
     
    
  
     
     
     
     
    
  
     
     
     
     
    
  
     
     [, , , , ]
     



     
     
     

Search Methods

Method	Trials	Best For	Command Flag
Grid	All combos	Small spaces (<20)	--search-method grid
Random	30-100	Large spaces	--search-method random
Bayesian	20-50	Expensive evals	--search-method bayesian

Grid Search

• Exhaustive coverage
• Good for discrete choices (layer sizes, termination penalty)
• Exponential cost with parameters — only practical for 2-3 params

Random Search

• Better than grid for >3 parameters
• Samples PPO and reward weights jointly from distributions
• Can set --hp-trials 50

Bayesian Optimization

• Builds surrogate model over the joint (PPO + reward weights) space
• Perturbs best-so-far config in both PPO and reward dimensions
• Best for expensive evaluations

How It Works

Each AutoML trial:

1. Sample PPO params   (lr, entropy, architecture, rollouts, ...)
2. Sample reward weights (position_tracking, termination, ...)
3. Write unified config JSON → starter_kit_log/automl/<id>/configs/
4. Launch train_one.py subprocess with both HP + reward overrides
5. Evaluate via TensorBoard logs
6. Score = f(reward_mean, success_rate, episode_length, termination_rate)
7. Update Bayesian model → suggest next trial

After all trials complete, the best (HP + reward weights) pair proceeds to full training.

Commands

# === PREFERRED: AutoML unified search ===
uv run starter_kit_schedule/scripts/automl.py `
    --mode stage `
    --budget-hours 12 `
    --hp-trials 8

# === SINGLE TRAINING RUN (manual HP selection) ===
uv run scripts/train.py --env <env-name>

# === ANALYZE RESULTS ===
uv run starter_kit_schedule/scripts/analyze.py `
    --metric episode_reward_mean `
    --sort descending

Analysis Output

# starter_kit_log/analysis/hyperparameter_importance.yaml
importance_ranking:
  # PPO params
  - parameter: learning_rate
    importance: 0.30
    best_value: 2.1e-4
  - parameter: entropy_loss_scale
    importance: 0.12
    best_value: 8.5e-4
  - parameter: policy_hidden_layer_sizes
    importance: 0.10
    best_value: [256, 128, 64]
  # Reward weights
  - parameter: termination
    importance: 0.18
    best_value: -200
  - parameter: approach_scale
    importance: 0.15
    best_value: 6.2
  - parameter: position_tracking
    importance: 0.08
    best_value: 2.5

top_5_configs:
  - config_id: config_023
    learning_rate: 2.1e-4
    entropy_loss_scale: 
     
     
     
    
    
     
     
     

Presets

Quick Validation (reward weights only)

# 6 configs — fix PPO defaults, explore reward scales
ppo_params: {}  # all defaults
reward_weights:
  termination: { type: "choice", values: [-500, -200, -50] }
  approach_scale: { type: "choice", values: [4, 8] }

Quick Validation (PPO only)

# 6 configs — fix reward defaults, explore PPO
ppo_params:
  learning_rate: { type: "choice", values: [1e-4, 3e-4, 1e-3] }
  rollouts: { type: "choice", values: [24, 48] }
reward_weights: {}  # all defaults

Full Unified Search

# ~200 configs (use random sampling)
# All PPO + network + reward weight parameters

Fine-Tuning Search

# Narrow ranges around best known config
ppo_params:
  learning_rate: { type: "loguniform", low: 1e-4, high: 5e-4 }
reward_weights:
  approach_scale: { type: "uniform", low: 4.0, high: 8.0 }
  termination: { type: "choice", values: [-300, -200, -100] }

Troubleshooting

Issue	Solution
All configs perform similarly	Widen search ranges (especially reward weights)
Best config unstable	Add constraint: high LR + many epochs
Search takes too long	Reduce --hp-trials, shorter training
Bayesian stuck	Increase initial random trials
Reward weights dominate	Fix reward weights, search PPO only first
PPO sensitive to reward scale	Search them jointly (default behavior)

Best Practices

1. Search PPO + rewards jointly — They interact; separate search is suboptimal
2. Start with random search — Better coverage than grid for the combined space
3. Use priors — Center search on known good values from cfg.py defaults
4. Add constraints — Filter unstable combinations early
5. Shorter training for search — 10M steps for comparison (5M too short to detect laziness)
6. Run multiple seeds — At least 3 per config for reliability
7. Log all configs — Stored automatically in starter_kit_log/experiments/
8. Diagnose first, then search — Use reward-penalty-engineering skill to identify WHAT to search before tuning weights
9. Review experiment history before launching — See training-pipeline Step 0
10. Watch for lazy robot at long horizons — 5M trials may look great but fail at 50M. Check that reached% is increasing, not just reward.

Empirical Findings

Task-specific empirical findings (best LR, rollout sweet spots, sprint-crash thresholds, deceleration moat, anti-laziness results, LR scheduler findings) are documented in: starter_kit_docs/{task-name}/Task_Reference.md → "Empirical AutoML Findings" section.

Always check this reference before launching new searches to avoid repeating experiments.

General lessons that apply across tasks:

Insight	Details
Joint PPO+reward search	Better than separate tuning — reward weights interact with LR/entropy
Trial length matters	5M steps too short to detect reward hacking; use ≥10M for serious candidates
Manual iteration wastes time	N manual train.py experiments should be 1 AutoML batch
Per-step bonuses are dangerous	Unconditional per-step rewards dominate goal bonuses at long horizons