In present scenario Internet has become an integral part of every one's life, as many services like mail, news, chat are available and huge amounts of information on almost any subject is available. However, in most cases the bandwidth to connect to the Internet is limited. It needs to be used efficiently and more importantly productively. Generally, bandwidth is distributed among groups of users based on some policy constraints. However, it turns out that the users do not always use the entire allocated bandwidth at all times. Also, some times they need more bandwidth than the bandwidth allocated to them. But when it is abundant then any kind of use can be permitted provided it is in consonance with the policy. The bandwidth usage patterns of users vary with time of the day, time of the year hence to stabilize this,a dynamic allocation of bandwidth that satisfies the requirements of the users is needed.

To maximize productive usage, a need to implement control access policies has to be implemented which prevents unproductive use but at the same time does not, to the extent possible, impose censorship. Squid proxy server is a an example of the same. Squid provides many mechanisms to set access control policies. However, deciding which policies to im¬plement requires experimentation and usage statistics that must be processed to obtain useful data. The proposed architecture elaborated in this paper is based on machine learning to de¬termine policies depending on the content of current URLs being visited. The main component in this architecture is the Squid Traffic Analyzer, which classifies the traffic and generates URL lists. The concept of delay priority will also be introduced which gives more options to system administrators in setting policies for bandwidth management.