FAQ about the Flatland Challenge¶

These are the most common questions regarding the Flatland Challenge. If your questions are not answered please check the Forum and post your question there.

How can I win prizes in this challenge?¶

You can win prizes in different categories.

Best Solution Prize: Won by the participants with the best performing submission on our test set. Only your rankings from the Round 1 and Round 2 are taken into account. Check the leader board on this site regularly for the latest information on your ranking.

The top three submissions in this category will be awarded the following cash prizes (in Swiss Francs):

CHF 7’500.- for first prize
CHF 5’000.- for second prize
CHF 2’500.- for third prize

Community Contributions Prize: Awarded to the person/group who makes the biggest contribution to the community - done through generating new observations and sharing them with the community.

The top submission in this category will be awarded the following cash prize (in Swiss Francs): CHF 5’000.-

In addition, we will hand-pick and award up to five (5) travel grants (up to 1’500 CHF each)to the Applied Machine Learning Days 2019 in Lausanne, Switzerland. Participants with promising solutions may be invited to present their solutions at SBB in Bern, Switzerland.

To check your eligibility please read the prizes section in the rules.

What are the deadlines for the flatland challenge?¶

The beta round starts on the 1st of July 2019 and ends on the 30th of July 2019
Round 1 closed on Sunday, 13th of October 2019, 12 PM. UTC +1
Round 2 closes on Sunday, 5th of January 2020, 12 PM. UTC +1

How is the score of a submission computed?¶

The scores of your submission are computed as follows:

Mean number of agents done, in other words how many agents reached their target in time.
Mean reward is just the mean of the cummulated reward.
If multiple participants have the same number of done agents we compute a “nomralized” reward as follows: .. code-block:
```
normalized_reward =cumulative_reward / (self.env._max_episode_steps +self.env.get_num_agents()
```
The mean number of agents done is the primary score value, only when it is tied to we use the “normalized” reward to determine the position on the leaderboard.

How do I submit to the Flatland Challenge?¶

Follow the instructions in the starter kit to get your first submission.

Can I use env variables with my controller?¶

Yes you can. You can access all environment variables as you please. We recommend you use a custom observation builder to do so as explained here.

What are the time limits for my submission?¶

If there is no action on the server for 10 minutes the submission will be cancelled and a time-out error wil be produced.

If the submissions in total takes longer than 8 hours a time-out will occur.

What are the parameters for the environments for the submission scoring?¶

The environments vary in size and number of agents as well as malfunction parameters. The upper limit of these variables for submissions are:

(x_dim, y_dim) <= (150, 150)
n_agents <= 250 (this might be updated)
malfunction rates this is currently being refactored

How can I experiment locally before submitting?¶

You can follow the instruction in the starter kit and use the provided example files to run your tests locally.

If you want to generate your own test instances to test your solution you can either head over to the torch baselines and get inspired by the setup there. Or you can generate your own test cases by using the same generators as used by the submission test set.

In order to generate the appropriate levels you need to import the malfunction_generator, rail_generator and schedule_generator as follows:

Then you can simply generate levels by instantiating:

stochastic_data = {'malfunction_rate': 8000,  # Rate of malfunction occurence of single agent
                   'min_duration': 15,  # Minimal duration of malfunction
                   'max_duration': 50  # Max duration of malfunction
                   }

# Custom observation builder without predictor
observation_builder = YourObservationBuilder()

width = 16 * 7  # With of map
height = 9 * 7  # Height of map
nr_trains = 50  # Number of trains that have an assigned task in the env
cities_in_map = 20  # Number of cities where agents can start or end
seed = 14  # Random seed
grid_distribution_of_cities = False  # Type of city distribution, if False cities are randomly placed
max_rails_between_cities = 2  # Max number of tracks allowed between cities. This is number of entry point to a city
max_rail_in_cities = 6  # Max number of parallel tracks within a city, representing a realistic trainstation

rail_generator = sparse_rail_generator(max_num_cities=cities_in_map,
                                       seed=seed,
                                       grid_mode=grid_distribution_of_cities,
                                       max_rails_between_cities=max_rails_between_cities,
                                       max_rails_in_city=max_rail_in_cities,
                                       )
# Different agent types (trains) with different speeds.
speed_ration_map = {1.: 0.25,  # Fast passenger train
                    1. / 2.: 0.25,  # Fast freight train
                    1. / 3.: 0.25,  # Slow commuter train
                    1. / 4.: 0.25}  # Slow freight train

# We can now initiate the schedule generator with the given speed profiles

schedule_generator = sparse_schedule_generator(speed_ration_map)

# Construct the enviornment with the given observation, generataors, predictors, and stochastic data
env = RailEnv(width=width,
              height=height,
              rail_generator=rail_generator,
              schedule_generator=schedule_generator,
              number_of_agents=nr_trains,
              obs_builder_object=observation_builder,
              malfunction_generator_and_process_data=malfunction_from_params(stochastic_data),
              remove_agents_at_target=True)

For the testing of you submission you should test different levels in these parameter ranges:

width and height between 20 and 150
nr_train between 50 and 200
n_cities between 2 and 35
max_rails_between_cities between 2 and 4
max_rail_in_city between 3 and 6
malfunction_rate between 500 and 4000
min_duration and max_duration in ranges from 20 to 80
speeds you can keep more or less equally distributed.

With these parameters you should get a good feeling of the test cases your algorithm will be tested against.

FAQ about the Flatland Repository¶

This section provides you with information about the most common questions around the Flatland repository. If your question is still not answered either reach out to the contacts listed on the repository directly or open an issue by following these guidlines.

How can I get started with Flatland?¶

Install Flatland by running pip install -U flatland-rl or directly from source by cloning the flatland repository and running python setup.py --install in the repository directory.

These Tutorials help you get a basic understanding of the flatland environment.

How do I train agents on Flatland?¶

Once you have installed Flatland, head over to the baselines repository to see how you can train your own reinforcement learning agent on Flatland.

Check out this tutorial to get a sense of how it works.

What is a observation builder and which should I use?¶

Observation builders give you the possibility to generate custom observations for your controller (reinfocement learning agent, optimization algorithm,…). The observation builder has access to all environment data and can perform any operations on them as long as they are not changed. This tutorial will give you a sense on how to use them.

What is a predictor and which one should I use?¶

Because railway traffic is limited to rails, many decisions that you have to take need to consider future situations and detect upcoming conflicts ahead of time. Therefore, flatland provides the possibility of predictors that predict where agents will be in the future. We provide a stock predictor that assumes each agent just travels along its shortest path. You can build more elaborate predictors and use them as part of your observation builder. You find more information here.

What information is available about each agent?¶

Each agent is an object and contains the following information:

initial_position = attrib(type=Tuple[int, int]): The initial position of an agent. This is where the agent will enter the environment. It is the start of the agent journey.
position = attrib(default=None, type=Optional[Tuple[int, int]]): This is the actual position of the agent. It is updated every step of the environment. Before the agent has entered the environment and after it leaves the environment it is set to None
direction = attrib(type=Grid4TransitionsEnum): This is the direction an agent is facing. The values for directions are North:0`,``East:1,South:2andWest:3`.
target = attrib(type=Tuple[int, int]): This is the target position the agent has to find and reach. Once the agent reaches this position its taks is done.
moving = attrib(default=False, type=bool): Because agents can have malfunctions or be stopped because their path is blocked we store the current state of an agent. If agent.moving == True the agent is currently advancing. If it is False the agent is either blocked or broken.
speed_data = attrib(default=Factory(lambda: dict({‘position_fraction’: 0.0, ‘speed’: 1.0, ‘transition_action_on_cellexit’: 0}))): This contains all the relevant information about the speed of an agent:
- The attribute 'position_fraction' indicates how far the agent has advanced within the cell. As soon as this value becomes larger than 1 the agent advances to the next cell as defined by 'transition_action_on_cellexit'.
- The attribute 'speed'' defines the travel speed of an agent. It can be any fraction smaller than 1.
- The attribute 'transition_action_on_cellexit' contains the information about the action that will be performed at the exit of the cell. Due to speeds smaller than 1. agents have to take several steps within a cell. We however only allow an action to be chosen at cell entry.
malfunction_data = attrib(default=Factory(lambda: dict({‘malfunction’: 0, ‘malfunction_rate’: 0, ‘next_malfunction’: 0, ‘nr_malfunctions’: 0,’moving_before_malfunction’: False}))): Contains all information relevant for agent malfunctions:
- The attribute 'malfunction indicates if the agent is currently broken. If the value is larger than 0 the agent is broken. The integer value represents the number of env.step() calls the agent will still be broken.
- The attribute 'next_malfunction' was REMOVED as it serves no purpose anymore, malfunctions are now generated by a poisson process.
- The attribute 'nr_malfunctions' is a counter that keeps track of the number of malfunctions a specific agent has had.
- The attribute 'moving_before_malfunction' is an internal parameter used to restart agents that were moving automatically after the malfunction is fixed.
status = attrib(default=RailAgentStatus.READY_TO_DEPART, type=RailAgentStatus): The status of the agent explains what the agent is currently doing. It can be in either one of these states:
- READY_TO_DEPART not in grid yet (position is None)
- ACTIVE in grid (position is not None), not done
- DONE in grid (position is not None), but done
- DONE_REMOVED removed from grid (position is None)

Can I use my own reward function?¶

Yes you can do reward shaping as you please. All information can be accessed directly in the env.

What are rail and schedule generators?¶

To generate environments for Flatland you need to provide a railway infrastructure (rail) and a set of tasks for each agent to complete (schedule).

What is the max number of timesteps per episode?¶

The maximum number of timesteps is max_time_steps = 4 * 2 * (env.width + env.height + 20)

What are malfunctions and what can i do to resolve them?¶

Malfunctions occur according to a Poisson process. The hinder an agent from performing its actions and update its position. While an agent is malfunctioning it is blocking the paths for other agents. There is nothing you can do to fix an agent, it will get fixed automatically as soon as agent.malfunction_data['malfunction'] == 0 . You can however adjust the other agent actions to avoid delay propagation within the railway network and keeping traffic as smooth as possible.

Can agents communication with each other?¶

There is no communitcation layer built into Flatland directly. You can however build a communication layer outside of the Flatland environment if necessary.

FAQ about bugs¶

Why are my trains drawn outside of the rails?¶

If you render your environment and the agents appear to be off the rail it is usually due to changes in the railway infrastructure. Make sure that you reset your renderer anytime the infrastructure changes by calling env_renderer.reset().

I keep getting there error when submitting from windows¶

When submitting from a Windows system you might run into the following error:

Please follow the intstruction in the starter-kit to avoid these problems. Make sure to reset your environment.yml correctly.

Link to Windows submission instructions