I state that I am creating a pacman reinforcement training. The idea is to encode the positions of objects throughout the map (walls, ghosts, cookies, powers, etc ...) in an n-dimensional tensor of 0 and 1 The situation is this, the game map is a 272x232 matrix: created with the points available:
x, y, w, h
Every object has these points, consequently, I encode everything in a tensor of 0 and 1 as follows:
def getWallMatrix(self):
matrix = np.zeros((self._h, self._w), dtype='uint8')
for w in self._actors:
if isinstance(w, Wall) or isinstance(w, Gate):
x, y, w, h = w.rect()
for i in range(0,w):
for j in range(0,h):
x = int(x)
y = int(y)
matrix[y+j-1][x+i-1]=1
return matrix
the initial idea was therefore to treat the tensor as an image and apply the function cv2.resize as follows:
observation[0] = cv2.resize(matrix, dsize=(RESIZE_H, RESIZE_W), interpolation=cv2.INTER_NEAREST)
With RESIZE_H = 68 and RESIZE_W = 68 the result is acceptable but the image is still too big for me, and if I had to put a lower value I would lose the resolution. This would create me some problems regarding the cookies distributed along the map, some disappear during the resizing.
I have seen around that someone has been able to acceptably resize the classic game map using only a 27x27 matrix of 0 and 1. How could I deal with the problem to have a compromise between very small image and quality?
