Conclusions and Future Work
9.2 Future Work
NN and fuzzy applications in ATM flow and congestion control were extensively studied and reported in the literature, as was shown in chapter 3. However, a number of issues still have to be looked at. In chapter 5 of this thesis, the ALP was examined in the flow control problem. In [ADAS96] the ALP was proposed for dynamic bandwidth allocation. An integration of both dynamic bandwidth allocation and flow control is possible using one ALP. The ALP scheme can declare its bandwidth requirements by looking at the past history of the traffic. If this bandwidth cannot be fully guaranteed then the traffic source is regulated by the difference between the guaranteed and the required (or predicted) bandwidth. Further investigation is required in this matter.
In the conventional threshold based congestion control scheme, congestion was detected by monitoring the violation of a certain threshold in the buffer. Chapter 6 has shown the greater ability of the NN to detect congestion when compared to the conventional threshold method. In the flow and congestion control schemes the
interest lays in detecting congestion in order to avoid cell loss. Bandwidth renegotiations for VBR video over ATM network have been proposed in [REIN96].
The scheme proposed in [REIN96] uses a dedicated (static) threshold in the buffer to determine the overflow and underflow regions and hence to determine the instances of renegotiations. It is believed that the NN proposed in chapter 6 can replace the threshold scheme in [REIN96] and can also provide better bandwidth utilisation. The NN scheme proposed in chapter 6 predicts the ratio of cell loss rate to the cell generation rate, the predicted output of the NN is then fed back to the source to reduce its rate by that amount. The same scheme can be used to renegotiate bandwidth by that explicit amount in order to avoid cell loss in real time video applications.
The static threshold was also used in other applications involving space priority and memory sharing in ATM switch [CHAUD94]. An analytical dynamic queue threshold scheme was proposed in [CHAUD96]. Using NN in such applications is worth investigation.
In this thesis, a multiplex of up to 8 sources was considered, however NN traffic predictors with a multiplex of 25 sources was used in [DROS96]. With the recent popularity of ATM, interest has risen in studying the behavior of TCP over ATM networks with some proposals to integrate TCP and ABR [SISA96, KALA95]. Clerot et al in [CLER98] has shown that NN traffic predictors can be used in dynamic bandwidth allocation of TCP traffic. One of the areas to be further investigated is the application of the intelligent technique to TCP traffic flow control.
If knowledge of the solution can be expressed in linguistic rules, a fuzzy system can be
networks can be used. Some of the limitations associated with neural networks and fuzzy logic were given in chapter 3. When using the FCC scheme, the rules and the membership functions were found by trial and error. To avoid the trial and error method that can be ineffective and unsatisfactory, especially if the scheme is to be implemented on-line, NN and Genetic Algorithms can be combined with fuzzy logic in hybrid systems such as Neuro-Fuzzy and Fuzzy-Genetic systems. So far the application of Neuro-Fuzzy in ATM CAC control [FONT96b] and congestion control [KWOK96, LIU96c] as well as the Fuzzy-Genetic approach in [RAMA96] have been used to optimise the membership functions (i.e. the fuzzy system parameters). In [NAUC97], Neuro-Fuzzy classifier, controller and function approximator were proposed. These Neuro-Fuzzy systems are able to “learn” linguistics rules (system structure) and membership functions (system parameters), or to optimise existing ones.
No work in ATM traffic control has been done using these systems. The use of these systems in ATM traffic control is important because the rule base (i.e. the system structure) can be found using supervised learning.
In [NAUC97] a Neuro-Fuzzy function approximator (NEFPROX) was tested as a time series predictor. The time series used was generated by the the Mackey-Glass equation [NAUC97]. The NEFPROX was compared with ANFIS [JANG93], a well-known function approximator. ANFIS gives a better approximation of the function, but due to a complex learning algorithm it takes a longer time to obtain this result. NEFPROX learning is very fast, but has a higher approximation error.
In appendix B of this thesis, NEFPROX was used as a congestion controller instead of the NN scheme described in chapter 6. The same training files, created in chapter-6, were used to train NEFPROX. The MSE learning curve, the membership functions
and rules generated by the NEFPROX are shown in appendix B. Appendix B shows that the MSE and number of rules generated is very high which makes the NEFPROX difficult to implement in the congestion control problem. However, further work (which is beyond the scope of this thesis) is needed to reduce the complexity (i.e. the number of rules) using pruning techniques [NAUC97].
Because of the introduction of different traffic classes such as CBR, real-time VBR, non-real-time VBR, ABR and UBR, which require different QoS guarantees such as CLR, CTD, CDV etc, the problem of CAC becomes more complex as the number of classes in the switch increases. Because of this complexity it would be very difficult for the stochastic or analytical algorithms to find a solution. A more detailed description of CAC in a time prioritised switch can be found in Appendix C.
Implementing the Neuro-Fuzzy classifier NEFCLASS proposed in [NAUC97] to the problem of CAC may be very interesting, since the rules and the membership functions will make it easier for the user to visualise how the different classes affect each other.
This thesis discussed the problem of ATM traffic control with time priorities assigned in the ATM switch however, the control problem becomes more complex if time and space priorities to be used in the same switch. Due to the complexity of the problem it is believed that machine intelligence can be a good candidate for solving the problem.