simulation_config.yml

# configuration file for the actor-critic PPM
# DESCRIPTION -----------------------------------------------------------------
# - metabolism: a dict introduced with the GridOrientedPPM Model. It contains
#       a dict for each species in the model (in the same order as in
#       densities). The dict contents are: {Amount of food_reserve reduced
#       per timestep, amount of fr gained from eating, amount of fr lost due to
#       giving birth}
Model:
    dim:                  !!python/tuple [10, 10]  # dimensionality
    densities:            !!python/tuple [0.01, 0.01]  # Predator, Prey
    food_reserve:         0
    max_food_reserve:     8
    metabolism:  # see description!
      OrientedPredator:
        fast:           0.25  # decrease every timestep
        satiety:        3  # get for eating
        exhaust:        3  # for mating
      OrientedPrey:
        fast:           1
        satiety:        2
        exhaust:        3
    generation:           0
    neighbourhood:        49  # only squares of odd numbers, e.g. 9, 25, 49
    p_breed:              0.4
    p_flee:               0.4
    p_eat:                0.6  # = 1-p_flee
    mortality:            False  # if False, agents are immortal, can't starve
    instadeath:           0.005  # for predators; if there is more than 1 predator, each predator can die with that probability
    rewards:
      wrong_action:       -1
      default_prey:       1
      default_predator:   1
      indifferent:        0
      succesful_predator: 5
      offspring:          20
      death_starvation:   -5
      death_prey:         -5
      default:            1
      instadeath:         0
Sim:
    goal:               "training"  # or "testing"; trainig is regular usage, testing is without using memory/history and without optimization
    steps:              1000 # training with more than 1000 timesteps reeeeeeeally slows down the optimization
    episodes:           10000
    save_state_to:      "plots/newtest/"  # store simulation
    resume_state_from:  ""  # resume but is also command line option
    record_values:      "generation, reward"
    save_state_every:   2  # episodes

Plot:
    every:              100
    render:             False
    filepath:           "plots/newtest/"
    figsize:            !!python/tuple [9,12]
    fmt:                "png"
    dpi:                150

Network:
    kind:               'fc'  # fully connected without any conv layers
    mode:               'cpu'  # can also be 'gpu'
    layers:             # I still need a convenient way to describe this
      input:            !!python/tuple [50, 40]  # Model -> neighbourhood + 1
      hidden1:          !!python/tuple [40, 40]
      hidden2:          !!python/tuple [40, 40]
      hidden3:          !!python/tuple [40, 40]
      action_head:      !!python/tuple [40, 27]
      value_head:       !!python/tuple [40, 1]
    gamma:              0.9  # discount factor