Training data from Reinforced SciNet

Summary:

The results from the training of neural networks in v2 of Reinforced SciNet, published partially in v2 of the paper Operationally meaningful representations of physical systems in neural networks.

File description:

results.txt - The results from the training during reinforcement learning.

results_loss.txt - The loss from the training during representation learning.

selection.txt - The noise level of latent neurons during representation learning.

Parameters: Reinforcement Learning

Server parameters

• 21 workers, 2 predictors, 1 trainer each
• 3M episodes

Training parameters

• glow: 0.1
• gamma: 0.01
• softmax: 0.5
• learning rate: 0.00005
• reward clipping: 1.0e-7

Network parameters

• DPS model:
  {'env1': [128, 128, 128, 128, 64, 32],
  'env2': [128, 128, 128, 128, 64, 32],
  'env3': [128, 128, 128, 128, 64, 32]}

Parameters: Representation Learning

Server parameters

• 21 workers, 2 predictors, 1 trainer each
• 5M episodes

Training parameters

• learning rate: 0.0001
• reward clipping: 1.0e-7
• selection discount: 0.04
• minimization discount: 0.02
• ae discount: 10.0
• agent discount: 1.
• reward rescaling: 10
• predicted actions: 1
• training data: 200K

Network parameters

• Prediction model:
  {'env1': [64, 128, 128, 128, 128, 64, 32],
  'env2': [64, 128, 128, 128, 128, 64, 32],
  'env3': [64, 128, 128, 128, 128, 64, 32]}
• Encoder model: [128, 128, 64, 32]
• Decoder model: [32, 64, 128, 128, 128]