Table 2.

Results. While our proposed approach of IDVQ + DRSC + XNES achieves not much lower scores (sometimes better; see Table 1), it does so using up to two orders of magnitude fewer neurons, and no hidden layers. The proposed feature extraction algorithm IDVQ+DRSC is simple enough (using basic, linear operations) to be arguably unable to contribute to the decision making process in a sensible manner (see Section 3.2.4). This implies that the tiny network trained on decision making alone is of sufficient complexity to learn a successful policy, potentially prompting for reconsidering the actual complexity of this standard benchmark. The number of neurons used in our approach solely depends on the size of each game actions space (see Table 1 for reference). The number of weights in our approach are scaled to the worst case of 150 centroids (1k for 6 neurons, 3k for 18 neurons), which in our runs was only reached by a few hard games: averages were more commonly in the 300 to 1k weights range

	HyperNeat	OpenAI ES	GA (1B)	NSRA-ES	Ours
# neurons	$\sim$ 3034	$\sim$ 650	$\sim$ 650	$\sim$ 650	6 – 18
# hidden layers	2	3	3	3	0
# connections	$\sim$ 906k	$\sim$ 436k	$\sim$ 436k	$\sim$ 436k	1k – 3k