hidden_units: 128
lambd: 0.95
learning_rate: 5.0e-3
max_steps: 200
max_steps: 300
memory_size: 256
normalize: false
num_epoch: 3