Add option to deactivate departure and arrival windows

[mmarti-tsch]

No description provided.

[mmarti-tsch]

a momentary fix works by:

def deactivate_windows(env):
for agent in env.agents:
agent.earliest_departure = 0
agent.latest_arrival = np.inf
agent.status = TrainState.READY_TO_DEPART

usage:
env = ....
env.reset()
deactivate_windows(env)

[mmarti-tsch]

A cleaner solution would involve sth. like a parameter:
env = RailEnv(...., include_windows=False)

and some internal logic which does not involve the arrival and departure times, and starts the agent's states in TrainState.READY_TO_DEPART

[mmarti-tsch]

I just found out that this is not enough to reinstate the previous behaviour of flatlatland (version 2).

reason: there is a check
if agent.state == TrainState.DONE and agent.arrival_time is None:
....

in the handle_done_state() method of RailEnv, where agents are removed from the environment.
It seems that on the original creation of an environment and first reset, the agent.arrival_time is always None.
Once the environment was run and agent reached target, agent.arrival_time is set to some integer value.
Upon another reset, it maintains this value...
Let's think about how to handle this.

[mmarti-tsch]

I found out that this is not so much an issue of the reset() but of "broken references".
E.g. if we do
env = RailEnv(...)
env.reset(...)
agent = env.agents[0]
now agent.arrival_time is None.

(bring agent to his target)
now
agent.arrival_time has some value, say 3.

if we do:
env.reset(...)
we now still have
agent.arrival_time = 3

but
env.agents[0].arrival_time is None.

So the issue lies in that our reference "agent" does not refer to env.agents[0] anymore.

My conclusion is that this is not really a bug.
It might still be a possible pitfall for users.

[aiAdrian]

The issue can be closed, right? mmarti-tsch ***@***.***> schrieb am Mi., 12. Juli 2023, 16:26:

…

I found out that this is not so much an issue of the reset() but of "broken references". E.g. if we do env = RailEnv(...) env.reset(...) agent = env.agents[0] now agent.arrival_time is None. (bring agent to his target) now agent.arrival_time has some value, say 3. if we do: env.reset(...) we now still have agent.arrival_time = 3 but env.agents[0].arrival_time is None. So the issue lies in that our reference "agent" does not refer to env.agents[0] anymore. My conclusion is that this is not really a bug. It might still be a possible pitfall for users. — Reply to this email directly, view it on GitHub <#10 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACPQ2XS3LJJPNLSYLMLOM43XP2XX3ANCNFSM6AAAAAAXXFEAKM> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>