flexible_marl_machine.yml

kind: 'rl'
parameters:

  model:
    kind: 'deep_rule'
    parameters:
      __rules__: ''

      policy:
        kind: 'flexible_action'
        parameters:
          policy_method: 'independent'

          model:
            __model__: ''

          action_selector:
            kind: 'phase_selector'
            parameters:
              default:
                kind: 'epsilon_greedy'
                parameters:
                  epsilon: 0.05
              phases:
                - phase:
                    kind: 'warm_up'
                    parameters:
                      step: 0
                  action_selector:
                    kind: 'uniform'
                - phase:
                    kind: 'warm_up'
                    parameters:
                      step: 1
                  action_selector:
                    kind: 'epsilon_greedy'
                    parameters:
                      epsilon: 0.4
                - phase:
                    kind: 'training'
                  action_selector:
                    kind: 'epsilon_greedy'
                    parameters:
                      epsilon: 0.4
                      min_epsilon: 0.05
                      decay_factor: 0.995
                      decay_steps: 40


  __encoder__: ''

  trainer:
    kind: 'dqn'
    parameters:
      decay: 1.0
      update_steps: 20
      epochs: 1

      memory:
        kind: 'replay'
        parameters:
          size: 1024
          batch_size: 128
          prefetch: 8

      loss:
        kind: 'mse'

      optimizer:
        model:
          kind: 'adam'
          parameters:
            lr: 0.001

      return:
        kind: 'no'
        parameters:
          discount_factor: 0.99