data/derived/chapter_hp_mapping.md

Chapter-to-HP Mapping

For each chapter, identify which config fields have >1 unique value (swept) vs exactly 1 value (fixed).

Chap1 (8100 runs)

Swept HPs:

• num_envs: 5 levels — [1024, 128, 256, 512, 64]
• rollout_length: 5 levels — [1024, 128, 256, 512, 64]
• clip_epsilon: 3 levels — [0.01, 0.2, 0.4]
• loss_weights.critic_loss: 3 levels — [0.5, 0.85, 1.0]
• loss_weights.actor_entropy: 3 levels — [-0.001, -0.01, -0.1]

Fixed HPs:

• discount = 0.99
• gae_lambda = 0.95
• gae_horizon = 0
• loss_weights.actor_loss = 1
• optimizer.lr = 0.0003
• optimizer.grad_clip_norm = 1.0
• normalize_obs = True
• normalize_gae = True
• minibatch_size = 128
• reuse_rollout_epochs = 4
• agent_network.actor_hidden_layer_sizes = [256, 256]
• agent_network.critic_hidden_layer_sizes = [256, 256]

Chap2 (8100 runs)

Swept HPs:

• num_envs: 5 levels — [1024, 128, 256, 512, 64]
• rollout_length: 5 levels — [1024, 128, 256, 512, 64]
• discount: 3 levels — [0.99, 0.995, 0.999]
• gae_lambda: 3 levels — [0.7, 0.8, 0.95]
• gae_horizon: 3 levels — [0, 32, 64]

Fixed HPs:

• clip_epsilon = 0.2
• loss_weights.actor_loss = 1
• loss_weights.critic_loss = 0.5
• loss_weights.actor_entropy = -0.01
• optimizer.lr = 0.0003
• optimizer.grad_clip_norm = 1.0
• normalize_obs = True
• normalize_gae = True
• minibatch_size = 128
• reuse_rollout_epochs = 4
• agent_network.actor_hidden_layer_sizes = [256, 256]
• agent_network.critic_hidden_layer_sizes = [256, 256]

Chap3 (4800 runs)

Swept HPs:

• num_envs: 5 levels — [1024, 128, 256, 512, 64]
• rollout_length: 5 levels — [1024, 128, 256, 512, 64]
• optimizer.grad_clip_norm: 4 levels — [0.0, 0.5, 1.0, 5.0]
• normalize_obs: 2 levels — [False, True]
• normalize_gae: 2 levels — [False, True]

Fixed HPs:

• discount = 0.99
• gae_lambda = 0.95
• gae_horizon = 0
• clip_epsilon = 0.2
• loss_weights.actor_loss = 1
• loss_weights.critic_loss = 0.5
• loss_weights.actor_entropy = -0.01
• optimizer.lr = 0.0003
• minibatch_size = 128
• reuse_rollout_epochs = 4
• agent_network.actor_hidden_layer_sizes = [256, 256]
• agent_network.critic_hidden_layer_sizes = [256, 256]

Chap4 (8100 runs)

Swept HPs:

• num_envs: 5 levels — [1024, 128, 256, 512, 64]
• rollout_length: 5 levels — [1024, 128, 256, 512, 64]
• optimizer.lr: 3 levels — [0.0003, 0.003, 3e-05]
• minibatch_size: 3 levels — [1024, 128, 512]
• reuse_rollout_epochs: 3 levels — [1, 4, 8]

Fixed HPs:

• discount = 0.99
• gae_lambda = 0.95
• gae_horizon = 0
• clip_epsilon = 0.2
• loss_weights.actor_loss = 1
• loss_weights.critic_loss = 0.5
• loss_weights.actor_entropy = -0.01
• optimizer.grad_clip_norm = 1.0
• normalize_obs = True
• normalize_gae = True
• agent_network.actor_hidden_layer_sizes = [256, 256]
• agent_network.critic_hidden_layer_sizes = [256, 256]

Chap5 (8100 runs)

Swept HPs:

• num_envs: 5 levels — [1024, 128, 256, 512, 64]
• rollout_length: 5 levels — [1024, 128, 256, 512, 64]
• optimizer.lr: 3 levels — [0.0003, 0.003, 3e-05]
• agent_network.actor_hidden_layer_sizes: 3 levels — ['[256, 256, 256]', '[256, 256]', '[512, 512]']
• agent_network.critic_hidden_layer_sizes: 3 levels — ['[256, 256, 256]', '[256, 256]', '[512, 512]']

Fixed HPs:

• discount = 0.99
• gae_lambda = 0.95
• gae_horizon = 0
• clip_epsilon = 0.2
• loss_weights.actor_loss = 1
• loss_weights.critic_loss = 0.5
• loss_weights.actor_entropy = -0.01
• optimizer.grad_clip_norm = 1.0
• normalize_obs = True
• normalize_gae = True
• minibatch_size = 128
• reuse_rollout_epochs = 4