Apparatus and method for learning current network behavior and predicting future behavior which utilizes a state transition graph. The graph includes nodes which represent network states, and arcs which represent trends in observable network parameters that result in a transition from a current state to another state. For example, a watch service may be instituted on multiple ports of a router, and the observed network traffic on the ports over time may be transformed into a state transition graph that represents network behavior. The network states may be labeled such as "good", or "bad", etc., according to a predetermined performance criteria. Once a state transition graph is constructed, the system may then monitor the current state and current trends of the network parameters in order to predict and display future network states. The system may include an automatic warning signal for alerting a user that the network is headed in the direction of a problematic state. The prediction of future network behavior may be made from the state transition graph, the current state of the network, and the current network trends.