Much of the current research involving the delivery of multimedia data focuses on the need to maintain a constant Quality of Service (QoS) throughout the lifetime of the connection. Delivery of a constant QoS requires that a guaranteed bandwidth is available for the entire connection. Techniques, such as resource reservation, are able to provide for this. These approaches work well across networks that are fairly homogeneous, and which have sufficient resources to sustain the guarantees, but are not currently viable over either heterogeneous or unreliable networks. To cater for the great number of networks (including the Internet) which do not conform to the ideal conditions required by constant Quality of Service mechanisms, this thesis proposes a different approach, that of dynamically adjusting the QoS in response to changing network conditions. Instead of optimizing the Quality of Service, the approach used in this thesis seeks to ensure the delivery of the information, at the best possible quality, as determined by the carrying ability of the poorest segment in the network link. To illustrate and examine this model, a service-adaptive system is described, which allows for the streaming of multimedia audio data across a network using the RealTime Transport Protocol. This application continually adjusts its service requests in response to the current network conditions. A client/server model is outlined whereby the server attempts to provide scalable media content, in this case audio data, to a client at the highest possible Quality of Service. The thesis presents and evaluates a number of renegotiation methods for adjusting the Quality of Service between the client and server. An A djusted QoS renegotiation method algorithm is suggested, which delivers the best possible quality, within an acceptable loss boundary.

Generative Multimedia (GM) applications are an increasingly popular way to implement interactive media performances. Our contributions include creating a metric for evaluating Generative Multimedia performances, designing a model for accepting perceiver preferences, and using those preferences to adapt GM performances. The metric used is imprecision, which is the ratio of the actual computation time of a GM element to the computation time of a complete version of that GM element. By taking a perceiver's preferences into account when making adaptation decisions, applications can produce GM performances that meet soft real-time and resource constraints while allocating imprecision to the GM elements the perceiver least cares about. Compared to other approaches, perceiver-directed imprecision best allocates impreciseness while minimizing delay.