Put Time Horizon back into the default training config for BC
max_steps: 5.0e4
memory_size: 256
batches_per_epoch: 10
time_horizon: 64
num_epoch: 5
num_layers: 2
summary_freq: 1000