DQN for flow

[pengyuan-zhou]

Hi,

This is a branch to enable DQN (also by ray) for flow. Currently it supports single agent and multi-agent tests on traffic_light_grid scripts.

Other examples can be easily enabled by modifying corresponding example file, env file and test file, according to the modifications in this branch.

Cheers, Pengyuan

[pengyuan-zhou]

Ya I have zero ideas why the build fails, it works on my pc (without the unittest).

[eugenevinitsky]

Putting aside the build failure, this sets RLlibs DQN as the default option. We'd prefer you make it so that there's a choice of algorithm, of which DQN is one. You can look at train.py in the flow_maddpg branch for an example.

[pengyuan-zhou]

Putting aside the build failure, this sets RLlibs DQN as the default option. We'd prefer you make it so that there's a choice of algorithm, of which DQN is one. You can look at train.py in the flow_maddpg branch for an example.

Sure, I'll revise it. BTW in that train.py, is the TD3 in setup_exps_rllib same with the one in h-baseline?

[pengyuan-zhou]

Putting aside the build failure, this sets RLlibs DQN as the default option. We'd prefer you make it so that there's a choice of algorithm, of which DQN is one. You can look at train.py in the flow_maddpg branch for an example.

Hi Eugene, I've modified the code according to mappdg's fashion. Cheers.

[eugenevinitsky]

Well, not quite. The implementation of TD3 in h-baselines is not identical to the rllib one.

[eugenevinitsky]

This looks good to me minus comments! @pengyuan-zhou if the changes are made I will merge.

[pengyuan-zhou]

@AboudyKreidieh @kanaadp @kjang96

[eugenevinitsky]

Hey! I'm really sorry about the state of things here, we are working on getting a new maintainer in the next week or so so this doesn't happen again.