A Metropolis criterion based fuzzy Markov game flow controller has been designed for coping with congestion in high-speed networks and networked systems. For such networks the complete and accurate information is not easy to obtain in real time because of the uncertainties and highly time-varying time delays. A viable alternative is to employ the Q-learning, which is independent of a mathematical model and prior knowledge as well as it can be used in game-theoretic framework. It enables to learn the needed parameters from the operating network environment. The fuzzy Markov game offers a promising platform for robust control in the presence of external disturbances and unknown parameter variations that are bounded. The Metropolis criterion can cope with the balance between exploration and exploitation in action selecting. In a similar game-theoretic framework and by employing SOM-based virtual sensors one recent solution to intrusion detection system from the literature, which is compatible with our network flow control design, is highlighted and outlined too. Simulation experiments demonstrate the proposed controller can learn to take the best action in order to regulate source flows. Thus it can guarantee high throughput and low packet loss ratio while efficiently avoiding the congestion.