Configuration of HighwayEnvironment_Single_Base.py could be more accurate

[paolo-crisafulli]

The configuration of highway-env for this agent is:

env = gym.make("highway-fast-v0", render_mode='human')
env.configure({
    "lanes_count": 1,  # (4)
    "collision_reward": -1,
    "right_lane_reward": 0.0, # (0.1),
    "reward_speed_range": [0,40],
    "high_speed_reward": 1,
    "action": {
        "type": "DiscreteMetaAction",
        "target_speeds": [0,5,10,15,20,25,30,35,40]}
})

It could be improved by specifying:

"action": {
        "type": "DiscreteMetaAction",
        "target_speeds": [0,5,10,15,20,25,30,35,40]}
        "lateral": False
}

Indeed, for 100 test runs the actions counter goes from: actions: [83, 1535, 4, 569, 809] to actions: [1048, 1149, 803, 0, 0], indicating that lateral actions were actually chosen when trained with the initial configuration.

Sorry for the late report, it's something I've found and fixed recently.

By the way, thank you for the very stimulating simulation exercise!

[favu100]

Thank you very much for the information! Indeed, the RL agent should not execute lateral actions in the single-lane environment.