simulation_vs_epsilon.py

import time

import numpy as np
from concurrent.futures import ProcessPoolExecutor
import pandas as pd

import common as cmn
from simulation_vs_theta_voi_unaware import run_episode


if __name__ == '__main__':
    # Parse arguments, if any
    parallel, data_folder, _, overwrite = cmn.common_parser()

    ### MAIN SIMULATION PARAMETERS ###
    theta = cmn.std_theta
    epsilon_vec = cmn.epsilon_vec
    num_frame = cmn.num_frame

    # Check if files exist and load it if there
    prefix = 'vs_epsilon'
    data_shape = (len(epsilon_vec), len(cmn.results_label))
    avg_results, filename = cmn.check_data(data_shape, prefix, data_folder, overwrite_flag=overwrite)

    # Start evaluating
    for m, epsilon in enumerate(epsilon_vec):
        # Check if data is there
        if overwrite or np.any(np.isnan(avg_results[m])):

            args = (theta, epsilon, num_frame)

            start_time = time.time()
            if parallel:
                with ProcessPoolExecutor() as executor:
                    futures = [executor.submit(run_episode, ep, *args) for ep in range(cmn.E)]
                    results = [f.result() for f in futures]
            else:
                results = []
                for ep in range(cmn.E):
                    print(f'\tEpisode: {ep:02d}/{cmn.E - 1:02d}')
                    results.append(run_episode(ep, *args))

            # Average the results
            avg_results[m] = np.mean(np.array(results), axis=0)

            # Save data (doing it every time is redundant, but it is safe)
            avg_df = pd.DataFrame(avg_results, columns=cmn.results_label)
            avg_df.insert(0, 'epsilon', epsilon_vec)
            avg_df.to_csv(filename, index=False)

            # Print time
            elapsed = time.time() - start_time
            print(f"\t...done in {elapsed:.3f} seconds")

        else:
            print("\t...already done!")
            continue