Replay an agent starting at some timestep

[marota]

Is your feature request related to a problem? Please describe.

When an agent fails at some point, we'd like to replay the timesteps just before and see if we could find a solution to avoid failure. We would like to avoid replaying the whole scenario again with that agent, and for instance inspect manually the state, running some simulations of different actions at that timestep

Describe the solution you'd like

From the logs of the agent, rebuild the environment/observation at the desired timestep of a scenario.

Describe alternatives you've considered

A clear and concise description of any alternative solutions or features you've considered.

Additional context

This can be useful for debugging as well

[Tezirg]

Unfortunately, there is no easy solution for this. The power flow needs to be ran again, with the same actions to get a usable environment.

Assuming the environment at t=0 can be rebuilt; That is, all config is known, including seeds, backend and difficulty:
And, not taking into account a potential Opponent:

import grid2op
from grid2op.Episode import EpisodeData

def resume_timestep(env, agent_path, episode_name, resume_timestep):
    # Build episode to id hash table
    episodes_name_to_id = {}
    for i, ep in enumerate(env.chronics_handle.real_data.subpaths):
        episodes_name_to_id[os.path.basename(ep)] = i

    # Reload previous actions
    epr = EpisodeData.from_disk(agent_path, episode_name)

    # Set env to desired chronics
    env.set_id(episodes_name_to_id[episode_name])
    env.reset()

    # Replay actions up to desired timestep
    for i in range(resume_timestep - 1):
        if env.done:
            raise "Episode game over before resume_timestep: check params"
        o, r, d, i = env.step(epr.actions[i])

    return o,r,d,i

if __name__ == "__main__":
    dataset = "rte_case14_realistic"
    agent_path = "/tmp/agent-logs/"
    episode_name = "0001"
    resume_step = 10

    # Build env
    # Cannot rebuild env from logs.
    # So it needs to be a manual check, for Parameters / seeds / etc..
    env = grid2op.make(dataset)

    obs, reward, done, info = resume_timestep(env,
                                              agent_path,
                                              episode_name,
                                              resume_step)