Can you help me paraphrase this paragraph? plzzzz. thank you so much !!!!!

Learning from the surrounding environment is one of the first learning methods humans experience. Humans naturally start learning by interacting with their environment. RL is inspired by the psychological and neuroscientific perspectives on animal behaviour and of the mechanism by which agents can enhance their control of the environment [111, 112]. RL involves making an agent learn how to map situations to actions appropriately to achieve the highest rewards [112]. The agent does not have previous knowledge of which actions to implement but has to learn which actions produce the most rewards by attempting them through trial and error. The features ‘trial’ and ‘error’ are the main and unique features of RL. Thus, the agent continues to learn from its experience to increase its rewards. One of the recent successful RL methods is the deep Q network [111]. Extensions of deep Q networks have been suggested, including double Q-learning [209], continuous control with deep RL [210] and prioritised experience replay [211]. RL has been implemented to solve several IoT issues. Studies by [212, 213] proposed an anti-jamming scheme that is based on reinforcement learning for wideband autonomous cognitive radios (WACRs). In [212], information about sweeping jammer signal and unintentional interference was used to distinguish it from other WACRs; RL was used with this information to learn a sub-band selection policy accurately to evade the jammer signal and interference from other WACRs. Similarly, in [213], an RL method based on Q-learning was trained to effectively avoid jamming attacks sweeping over a wide spectrum of hundreds of MHz in real time. In the same direction, [214] used RL to develop an anti-jamming scheme for cognitive radios and integrate the scheme with deep CNN to improve the efficiency of RL in a large number of frequency channels. A similar scheme against aggressive jamming was proposed using deep RL in [215], in which jamming was considered activated in an aggressive environment, which is normally expected in tactical mobile networking; the results showed that RL is a promising method of developing schemes against aggressive jamming. V. IOT SECURITY LAYERS BASED ON ML AND DL METHODS In this section, we classify the previous studies on ML and DL methods for IoT security according to the layers these methods intend to protect. Even though ML and DL may be applied to protect more than one layer or the end-to-end system (which is the advantage of ML and DL methods over other methods and holds potential future uses), the following classification is proposed to highlight the conceptualisation of the ML and DL methods for IoT security. At the end of this section the technology tools that can essentially enable ML/DL deployment for IoT security are listed. A. Perception layer One of the promising applications of DL methods is physicallayer authentication. Traditional physical-layer authentication techniques apply assumption checks and relate the randomness and exclusiveness of the radio channel between “Alice” and “Bob”, to detect spoofing attacker “Eve” in a wireless network. Nonetheless, such an approach is not always practical, specifically in dynamic networks [216]. Wang, Jiang, Lv and Xiao [216] used a learning model to construct a physical-layer authentication model that uses past data generated from a spoofing model as learning vectors to train an extreme learning machine. The proposed model exhibited improved spoofing detection performance and consequently achieved considerably enhanced authentication accuracy compared with that of stateof-the-art methods. Shi, Liu, Liu and Chen [217] proved that the present Wi-Fi signals generated by IoT objects can be adopted to detect distinctive human behavioural and physiological features and can be utilised to authenticate individuals on the basis of an GANs The GAN framework simultaneously trains two models (i.e. generative and discriminative models) via an adversarial process. The generative model learns the data distribution and generates data samples, and the discriminative model predicts the possibility that a sample originates from the training dataset rather than the generative model (i.e. evaluates the instance for authenticity). In GANs, generating a sample needs only one pass through the model, unlike in DBNs and RBMs in which an unidentified number of iterations of a Markov chain is required [205, 207]. GAN training is unstable and difficult. Learning to generate discrete data by using GAN is a difficult task [205, 207]. GANs can be used to build an architecture for securing the cyberspace of IoT systems [206]. EDLNs EDLNs can be accomplished by merging generative, discriminative or hybrid models. Combining DL classifiers can help achieve model diversity, improve model performance and expand model generalisation. The time complexity of the system can be significantly increased. The use of GANs in securing IoT systems needs further investigation, particularly the possibility of implementing light homogenous or heterogeneous classifiers in a distributed environment to improve the accuracy and performance of a system. 24 understanding of their daily activities. The authors proposed a scheme which adopts a single pair of Wi-Fi signals generated by IoT devices to mine Wi-Fi channel state information and thus obtain the amplitude and the relative phase for precise user authentication without the need for user participation. Using these features, the authors developed a DL model (i.e. Deep Neural Network (DNN)) to identify the daily human activity distinctiveness of each individual and subsequently generate a fingerprint for each user, called Wi-Fi fingerprint, to capture the distinct characteristics of different users; the proposed DLbased authentication method exhibited high accuracy [217]. This study validates the potential application of DL algorithms in constructing authentication systems. In another study [215], a scheme against aggressive jamming was developed using RL, and jamming was considered activated in an aggressive environment, which is normally expected in tactical mobile networking. RL was found effective in developing a method against aggressive jamming [215]. The research in [218] also considered the issue of jamming in an IoT network and introduced a centralised approach to addressing possible jamming attacks in an IoT environment, which consists of resource-constrained devices. The idea of the proposed model is to use the IoT access point to protect against the jamming attacker by distributing its power over the subcarriers in an intelligent manner and using an evolutionarybased algorithm. The proposed method can converge in a practical iteration number; thus, it can provide a better solution than a random power allocation strategy. Along the same direction, two previous studies [212, 213] proposed RL-based anti-jamming schemes for WACRs. In [212], the authors used information about sweeping jammer signal and unintentional interference to distinguish it from those of other WACRs. This information and RL were combined to learn a sub-band selection policy accurately to evade the jammer signal and interference from other WACRs. Similarly, in [213], an RL method based on Q-learning was trained to effectively avoid jamming attacks sweeping over a wide spectrum of hundreds of MHz in real-time. In the same direction, [214] used RL to develop an anti-jamming scheme for cognitive radios and integrate it with deep CNN to improve the efficiency of RL in a large number of frequency channels. Incorporating cognitive radio (CR) capability into IoT devices has paved the way for an innovative research on IoT systems [219]. Currently, many researchers are conducting studies on communication and computing in IoT systems. According to two previous studies [220, 221], IoT systems cannot be sustained without comprehensive cognitive capability because of growing issues. CRs are radio devices that can learn and change in accordance with their dynamic environment [222]. The main step towards accomplishing such cognitive operation is enabling CRs to sense and understand their working environment. Ideally, CRs should be able to work over a wide frequency range. However, sensing all required frequencies in real time is a challenging task, specifically with the existence of jamming attacks. CRs can become increasingly useful and reliable communication systems if they can eliminate the incidence of accidental interference or deliberate jamming attacks.