04-07-2013, 04:43 PM
Survey on Genetic Algorithm and Artificial network
Survey on Genetic Algorithm.doc (Size: 71.5 KB / Downloads: 19)
Abstract:
As Security is censorious issue for computer networks, an intrusion detection system is one of the indispensable building blocks to make network secure and undeviating by misuse and abuse of computer system. Genetic search methods based on correlation based feature selection (CFS) was adapted to identify the ideal features for artificial immune system that work as the classifier which helps to classify threats and non threats efficiently over Network Intrusion Detection System (NIDS).
Introduction:
Security is the exemption from anxiety and intrusions are the activities that despoil the security policy and by attempt to compromise the integrity, confidentiality or availability of a resource [5]. So due to that speedy growth of intrusion in network the security concern has made increasingly important. Now IDS is the potential solution which extinguishes the restriction of traditional network security methods as virus detection, frangibility evaluation, and firewall. It allows the system to pay peculiar attention to dangerous components; to detect rapidly propagated viruses and to extenuate serious damage at an primal stage and when intrusion is detected, it can stop them at that point and limit it from succeeding [1][2].
IDS has mainly two types as Network-based intrusion detection systems (NIDS) is disseminate and autonomous in nature and works on the percept of signature matching. It monitors network, collect packets, filter and analyze traffic that passes through that network location by connecting software or hardware, resides in one or more systems connected to a network by network connecting devices. It has few advantages as it is an excellent way of detecting unauthorized users even before gaining access to the computer. It is desirable for medium to large scale organizations. They are resistless devices and it is not susceptible to direct attack and may not be identify by attackers.[5] Disadvantages over it as if a network uses encrypted packets or a high-speed network, then the NIDS technique may not be beneficial because the IDS does not have access to the keys of every devices in the network.
Reinforcement learning:
Reinforcement learning (which is true AI) is the formal computation model which describes how an animals and artificial systems can learn to optimize their behavior to obtain rewards and to avoid punishments. Its algorithms are closely related to methods of dynamic Programming for optimal control. It is more generally applicable than dynamic programming, since (i) it does not require sweeps over the entire state space and (ii) it does not depend on the assumption that the probabilistic nature of the environment as well as the reward structure are known.
Reinforcement learning is different from supervised learning, it is an important kind of learning, but alone it is not adequate for learning from interaction. In interactive problems it is often impractical to obtain examples of desired behavior that are both correct and representative of all the situations in which the agent has to act. In uncharted territory-- where one would expect learning to be most beneficial--an agent must be able to learn from its own experience.
Policy iteration
A policy is an assignment of actions to state. Policies can be stochastic, specifying the probability of performing each action at each state. The solution to the optimization problem of expression 1 can be cast as deterministic policy which is constant over time, so the agent performs the single same action every time it gets to a particular state. Policies without time dependence are called stationary. In policy iteration, a policy is first evaluated, and then improved. Policy evaluation consists of working out the value of every state x under policy π, ie the expected long term reward available starting from x and following policy π. The notation π(x) for the action that policy π recommends at state x.
Monte Carlo Control:
Policy iteration alternates between improving the current policy to arrive at a new policy, and then evaluating that new policy. To extend Monte Carlo evaluation to control, it suffices to insert improvement steps between the repeated evaluation steps. Monte Carlo control algorithm cannot converge to a suboptimal policy. If it were to do so, then the value function corresponding to that policy would eventually be learned (via Monte Carlo evaluation), at which point it would be determined that alternative actions are preferable. Convergence requires both the policy and the value function to be optimal.