configuration TorcsDQN {
    context : cpu

    learning_method : reinforcement
    environment : ros_interface {
        state_topic : "preprocessor_state"
        terminal_state_topic : "prepocessor_is_terminal"
        action_topic : "postprocessor_action"
        reset_topic : "torcs_reset"
    }

    reward_function : torcs.agent.dqn.reward

    num_episodes : 20000
    discount_factor : 0.999
    num_max_steps : 999999999
    training_interval : 1

    use_fix_target_network : true
    target_network_update_interval : 500

    snapshot_interval : 1000

    use_double_dqn : true

    loss : euclidean

    replay_memory : buffer{
        memory_size : 1000000
        sample_size : 32
    }

    strategy : epsgreedy{
        epsilon : 1.0
        min_epsilon : 0.01
        epsilon_decay_method: linear
        epsilon_decay : 0.0001
    }

    optimizer : rmsprop{
        learning_rate : 0.001
    }
}