SOLVED: The Q-learning approximation algorithm starts with an initial parameter estimate of 0. As the tabular Q-learning; upon observing data tuple (s, c, R(s, c), s') the target value y for the Q-value of (s, c) is defined as the sampled version of the Bellman operator, y = R(s, c) + rmaxQ(s', c, 0) Then the parameter theta is simply updated by taking a gradient step with respect to the squared loss L(theta) = sum(y - Q(s, c, theta))^2. The negative gradient can be computed as follows: (Enter y

What is Q-Learning? Beginners Guide