Computer Science Sample Assignment Part 3

Course: Computational Intelligence
Topic: Learning Automata
Level: Master of Computer Science

Paper1.pdf is the original paper that we received. In that paper they have proposed a solution for cognitive radio network. They proposed solutions containing two phases. In phase 1 channel and powers are allocated to base stations and in phase 2 each base station allocates channels to their subscribers.

The purpose of this assignment is we have to use one the LA algorithm to enhance their proposed solution.

Objective of the work:

You would need to solve the problem that you have chosen using a Learning Automata (LA) approach. A comprehensive tutorial on LA is attached (LA.pdf), in complement to the course notes.

What you are expected to achieve is:

Implement the proposed LA-based algorithm (which we use sela) and provide simulation results (Java/C++/C skills are required). So we already search and find out the best LA algorithm that fit in our problem to enhancing that paper is SELA (Stochastic Estimator Learning Automata).

You do this implementation in java and show some graphical result comparing the original answer and the our proposed solution phase, is examined. CAC aims at guaranteeing the QoS requirements by deciding whether a new call-request should be accepted or not. For this purpose, it takes into consideration the available network resources, the trac parameters and the QoS requirements of both the incoming call and the already established connections. CAC, which tries to avoid congestion while making as many connections as possible, should operate in real-time, being simple with no computational overhead or excessive storage requirements. The difficulty in providing an acceptable solution is shown by the problems that some of the most known CAC mechanisms of the literature encounter, a brief review of which will be presented in Section 2. Their drawbacks in combination with the dynamic nature of trac in ATM networks have led to the use of Arti®cial Intelligence techniques. In this paper, a new methodology which uses a simple Learning Automaton (LA), called Stochastic Estimator Learning Algorithm (SELA), is proposed in order to enhance the analytical approximations of the literature aiming at increasing the gain of the statistical multiplexing while guaranteeing the QoS requirements. The feedback that SELA receives is a function of the equivalent bandwidth approximation, the accuracy of which SELA attempts to improve. The suggested approach, as will be shown through a simulation study, remains impervious to trac variations in individual connections and achieves more statistical gain than the equivalent bandwidth approximations that it uses. The stochastic and nonlinear character of SELA maximises the number of the established connections while guaranteeing the QoS requirements. Finally, the proposed algorithm operates in real-time and can be implemented in hardware according to the procedures described in. This paper is organised as follows: First of all, in Section 2, related work of the literature is discussed, while, in Section 3, the proposed methodology is presented. In Section 4, a simulation comparative performance study is presented and ®nally, in Section 5, our conclusions and some issues for future work are discussed. 2. Related work Many di€erent CAC algorithms have been proposed in the literature trying to perform an e€ective admission control; an ecient classi®cation as well as a review of them have been presented in [3]. In summary, according to the methodology on which each algorithm is based, they can been classi®ed into equivalent bandwidth approximations [4±9,15,16,33,34], heavy trac approximations [17], upper bounds of the cell loss probability [18,19,32,35], mechanisms based on Arti®cial Intelligence techniques [10,11,20±22,28±31], Fast Resource Reservation and Time Windows mechanisms. The last two categories can be applied only to very speci®c situations and thus, they cannot be regarded as CAC algorithms suitable for general use. The ®rst three categories are based on mathematical computations that approximate the equivalent bandwidth or the expected cell loss probabilities and they perform CAC according to these estimations. Their drawback is that each of them is based on speci®c trac models and that they can be applied only to sources that comply with these models. However, the most signi®cant drawback is that they overestimate the equivalent bandwidth or the expected cell loss probability, which result in under-utilisation of the network resources. Moreover, their e€ectiveness depends considerably not only on the trac types of the sources, but also on the utilisation of the network resources, the QoS constraints, as well as on the statistical multiplexing that is achieved [6]. On the other hand, the mechanisms based on Arti®cial Intelligence techniques (mainly on Arti-®cial Neural Networks ± ANNs ± [10,11,20,22,28±31] or a combination of them with Fuzzy Logic[21]) do not depend on speci®c trac models. The ANNs are trained o€-line according to speci®c trac patterns and QoS requirements. However, their convergence to the correct decision (acceptance or rejection of the call-request) during their on-line operation depends heavily on whether the real trac characteristics resemble the ones that the ANN has been trained for. As the scenarios that an ANN should be trained for increase, so does its convergence time, its size and the 342 A.F. Atlasis et al. / Computer Networks 34 (2000) 341±353

B. The Two-Phase Resource Allocation Scheme

Phase 1 - Global Allocation:

In this phase, channels and transmit powers are allocated to BSs so that the interference caused to each PU is below a tolerable threshold. At the same time, we aim to cover as many CPEs as possible. When talking about coverage here, we do not care whether a CPE is active or idle. That will be taken care of in the second phase of the TPRA scheme. Consider a particular channel c. For each BS, the higher power it transmits on c, the more CPEs it can cover. However, the higher the transmit power of the BS, the more interference it causes to PUs and other cells. This interference reduces the number of CPEs that can be covered using channel c in other cells. We note that it is extremely hard to fully characterize the above dual effects of varying base stations transmit powers on the number of CPEs being covered in the whole network. Therefore, we rely on the following intuition for making our channel/power allocation decisions. A BS that is near any PU using channel c should only transmit at low power to reduce interference. On the other hand, a BS that is faraway from all PUs using channel c can transmit at higher power. Each BS can use a set of channels on which it can transmit at high power to cover faraway CPEs.