In this thesis work we study, through statistical physics methods, a problem coming from game theory: the dynamics of a set of learning agents with linear quadratic payoffs, in discrete time. In this model, the agents interact with each other over a network where couplings can be either positive or negative (correlated or anticorrelated); and the agents update their actions only based on their most recent payoff (and at most their last played action). The payoff itself is the only feedback or knowledge the agents have from the system, and depends as an aggregate on the action and coupling of each one of the agent's neighbours.