By using the following command, you can rollout your agent models obtained from training process above using problems in train set:
python main.py --rollout --agent_load_dir MyAgentLoadDir --agent_for_rollout MyAgent --optimizer_for_rollout MyOptimizer --log_dir MyLogDir
But before running it, please make sure that the 21 agent models named checkpointN.pkl saved from training process are in a folder named your agent class name MyAgent, and this folder is in directory MyAgentLoadDir, which seems like:
MyAgentLoadDir
│
└─ MyAgent
│
├─ checkpoint0.pkl
├─ checkpoint1.pkl
├─ ...
└─ checkpoint20.pkl
After rollout, in MyLogDir/rollout/runName or output/rollout/runName by default, MetaBox will generate a file named rollout.pkl which is a dictionary containing:
cost is the best costs sampled every 400 function evaluations along the rollout process of each checkpoint model running on each problem in train set.fes is the function evaluation times used by each checkpoint model running on each problem in train set.return is the total reward in the rollout process of each checkpoint model running on each problem in train set.