def step(self, action, outdir='test'):
"""
Action is a list of 6 inputs that is in the continous range [-1.0, 1.0]
This is constrained so that we can use a hyperbolic tangent output activation for our opolicy network.
This is particualarly helpful for stochastic policies (Gaussain) often used in SAC
"""
#TODO Clean this routine
#Translate action (list) to override params dict
#Translation is min + ([-1,1] --> [0,1]) * (mx - min)
mods = AttrDict()
mods.real_cycle_time= self.min_action[0] + (0.5 * (action[0]+1)) * (self.max_action[0] - self.min_action[0])
mods.vent_time_fract= self.min_action[1] + (0.5 * (action[1]+1)) * (self.max_action[1] - self.min_action[1])
#Hardcoded
mods.cycles=31
mods.real_vent_time=mods.vent_time_fract*mods.real_cycle_time
mods.input_orifice=2.78
mods.vent_orifice=1.56
mods.blowdown_orifice=3.5
#Simulate
data, ret, param, pickle_name,out_place= psa.simulate(mods,
outdir=outdir,
params_file='params',
verbose=False)
#TODO Verify if we care about prod_y or counter_y? Also [-1] refers to the last timestep but why [1]? We care about the middle cylinder?
reward = ret.prod_y[-1][1]
#New state
state = psa.init(self.params, self.norm)
return state, reward, True, {}
def reset(self):
state = psa.init(self.params, self.norm)
return state