We propose a new way of characterizing the complexity of online problems.Instead of measuring the degradation of the output quality caused by the ignoranceof the future we choose to quantify the amount of additional global informationneeded for an online algorithm to solve the problem optimally. In our model, thealgorithm cooperates with an oracle that can see the whole input. We definethe advice complexity of the problem to be the minimal number of bits(normalized per input request, and minimized over all algorithm-oracle pairs)communicated by the algorithm to the oracle in order to solve the problemoptimally. Hence, the advice complexity measures the amount of problem-relevantinformation contained in the input. We introduce two modes of communication between the algorithm and the oraclebased on whether the oracle offers an advice spontaneously (helper) or onrequest (answerer). We analyze the Paging and DiffServ problems in terms ofadvice complexity and deliver upper and lower bounds in both communication modes;in the case of DiffServ problem in helper mode the bounds are tight.
