karinemiras / evoman_framework Goto Github PK

ElementTree.getchildren seems to be deprecated:

Can be fixed by:

for c in list (tag):

Hi, I'm debugging my assignment 2 code, and I find this issue. Maybe you would want to take a look at it?

In demo_controller.py, the control method does a min-max scaling:

def control(self, inputs, controller):
    # Normalises the input using min-max scaling
    inputs = (inputs-min(inputs))/float((max(inputs)-min(inputs)))

The problem is that min-max scaling should be done on the entire data set, not on individual inputs, otherwise, it will result in distorted data ranges.

In this case, this scaler will completely shadow the inputs[2] and input[3] because the original value range of these two is [-1, 1], while all other value ranges are wider than [-400, 400] based on my test. This means that after scaling, the original value of inputs[2] and inputs[3] won't matter because it is not significantly big enough compared to other values. This is bad because these two inputs represent "player's direction" and "enemy's direction", but now these two values in this case can not be recognized by NN.

The following graph is an illustration. I collected some data from the sensor to draw this figure, the y-axis is the scaled inputs[2], and the x-axis is calculated without original inputs[2] and inputs[3]. They show a directly proportional relationship with k=1, meaning that they are basically the same and the scaled inputs[2] is irrelevant with inputs[2] from the sensor.

I'll suggest to change it to something like this, which will make the scaling more resonable:

def control(self, inputs, controller):
    # inputs_range represents real range for each inputs from sensor
    inputs_range = np.array([[-700, 700], [-400, 400], [-1.0, 1.0], [-1.0, 1.0], [-700, 700], [-400, 400],
                             [-700, 700], [-400, 400], [-700, 700], [-400, 400], [-700, 700], [-400, 400],
                             [-700, 700], [-400, 400], [-700, 700], [-400, 400], [-700, 700], [-400, 400],
                             [-700, 700], [-400, 400]])
    inputs = (inputs-inputs_range[:,0])/float((inputs_range[:,1]-inputs_range[:,0]))

Maybe you would want to fix this after the assignment 2 finished? Since lots of groups already working on their project based on this NN structure.

From the documentation: "
play(self, pcont=None, econt=None): executes a game run using the solu-
tions given as ’pcont’(player) and ’econt’(enemy) for the controllers. When
using the default controllers these parameters are not necessary, as agents
will take random actions.
When in single objective mode, it returns: fitness, player life, enemy life,
and game run time, in this order as single values. When in multi objective
mode, it returns: fitness, player life, enemy life, and game run time, in this
order as 4 lists containing values of each game (lists are ordered by the order
of enemies parameter provided)."

But actually in multimode is returning the mean - std of the player life, enemy life and time instead of the arrays