Proc. of the 16th Workshop “From Object to Agents” (WOA15) June 17-19, Naples, Italy Mobile Agents with Recurrent Neural Networks-based Computing Model for Echo Cancellation Problem F. Bonanno G. Capizzi G. Lo Sciuto Department of Electrical, Electronics and Department of Electrical, Electronics and Department of Engineering Informatics Engineering Informatics Engineering Roma Tre University University of Catania University of Catania Rome, Italy Catania, Italy Catania, Italy Abstract—The purpose of this paper is to draw attention to a novel solving method for AEC problem and to search for a solution based on mobile agents (MAs) technology. Several applications benefit already by the use of MAs technology and we now want to join a soft computing technology as recurrent neural networks (RNNs) for this problem’s area in communication systems. In this paper we propose a MAs with RNNs-based, so as pipelined recurrent neural networks (PRNNs), computing model Fig. 1. A problem in Acoustic Echo Cancellation: Line/Network Echo for AEC’s problem in communication systems. Prediction of echo Cancellation. paths can be performed by RNNs and PRNNs based processing. The results about the faced problem in communication systems, where the use of MAs with RNNs implementation might be of an agent network. an improvement over a more conventional solutions, are here Several are the existing available agent platforms including summarized, presented and evaluated. Satisfactory performance (MOLE, Aglets, Concordia, Ara, TACOMA, and Mobile- in echo cancellation were obtained. The two echo signals input, as echo remote and echo local are provided as inputs to the RNN and C is an embeddable mobile agent system compliant with PRNN and the echo output prediction shows a relevant reduction Foundation for Intelligent Physical Agents (FIPA) that is an evaluated as 60 db. internationally recognized agent standards. We have implemented common RNNs and PRNNs for I. I NTRODUCTION adaptive filtering but in recent studies appeared in literature, to reduce the computational complexity of the bilinear recurrent Several are the application of adaptive filtering algorithms neural network (BLRNN), a novel low-complexity nonlinear in communication systems problems as for speech recognition, adaptive filter with a pipelined bilinear recurrent neural net- echo cancellation, interference suppression, noise cancellation work (PBLRNN) was presented by some researchers. and acoustic echo cancellation. There are also two main The present paper in this area includes development and categories of audio analysis problems area that could be implementation of new computing model for adaptive filtering investigated as sound matching and speech recognition. by using MAs in conjunction with RNNs. MAs is one promis- Real-time speech recognition is an important task in current ing new paradigms for distributed application. A mobile agent digital communication systems such as mobile telephone sys- consists of the program code and the program execution state. tems, adaptive filtering, algorithm for echo cancellation etc... Initially, a mobile agent resides on a computer called home In this paper we deal with the use of mobile agents (MAs) machine or dispatching server. The agent is then dispatched technology with adaptive filtering for thefor echo cancellation to execute on a remote computer called mobile agent host. problem. The agent concept has been widely adopted in When a mobile agent is dispatched, its entire code and the many areas such as: control system, network management, execution state are transferred to the mobile agent host. The information management, E-commerce. host provides a suitable execution environment for the mobile A mobile agent is the composition of computer program agent. and data which can travel from one computing platform to another. The agent technology becomes popular for the reasons such as: parallel performance of tasks, dynamic adaptation A. The Acoustic Echo Cancellation Problem to changing conditions, easy deployment of new program and being able to exchange information. In a mobile agent Echo Cancellation is used to enhance speech for Radio, network, agents can carry data and programs while moving Mobile, VoIP applications are available for echo cancellation from one computing platform to another and one task can solutions including acoustic echo cancellers (AEC) and line or be decomposed into several sub-tasks. These agents work network echo cancellers. AEC is an essential part for providing cooperately and dynamically adapt themselves to the changing voice quality enhancement in telephone communications (see environment and as known scalability is one important feature Fig. 1). 109 Proc. of the 16th Workshop “From Object to Agents” (WOA15) June 17-19, Naples, Italy In drive-thru order posts, reflections off of curbs and other various buildings and structures create an unique acoustic environment to every deployment. The impulse response (echo path) will be very dissimilar to that of a typical office envi- ronment and potentially could have a long echo tail. Besides the uniqueness of the echo path, a drive-thru application also has to be able to handle the non-stationary aspects of the background noise. This makes a noise reduction algorithm a requirement with AEC. There are also applications in which the loudspeaker and microphone are not physically tied to the same device, as in distributed multimedia systems. This reduces the acoustic coupling between loudspeaker and micro- phone. In this situation assuring synchronization between the Fig. 2. Schematic of a proposed echo cancellation system based on MAs sampling rates of the loudspeaker and microphone becomes an with RNNs and a main Server RNN0 additional burden. At present, most teleconferencing systems involve a single Cancellation is the reduction of the reflected copies of full-duplex audio channel for voice communication. These a direct path wave in a signal. An AEC operates on the systems usually employ an acoustic echo canceler to re- digitally sampled audio signals of the communication device. move undesired echoes that result from coupling between The transfer function of the acoustic environment from the the loudspeaker and microphone. As these systems evolve to loudspeaker to the microphone on the device is estimated to transparent audio-video medium, the need for enhanced sound cancel the received echoes from the microphone signal then realism becomes more important. an AEC is required. This need leads to consideration of multichannel audio, An adaptive filter is conventionally used in voice echo which at a minimum involves two channels, i.e., stereophonic cancellation to accommodate the time varying nature of the sound. However, before full-duplex stereophonic teleconfer- echo path. The filter learns the path when the far-end speaker encing can be deployed, the AEC problem must be solved. is talking and the near-end speaker is silent and adjusts its In this section we have reported these considerations in coefficients according to the algorithm optimization. For an order to discuss preliminary some unsuccessful attempts in adaptive filter to learn the echo path it must have an undis- AEC problems area. Several applications benefit already by turbed reference signal to adapt to. Unfortunately in double- the use of mobile agents technology and we now want to talk detection this filtering scenario cannot be admitted as the use a soft computing technology as RNNs for this problem near-end speaker may want to interrupt the far-end speaker. area in communication systems. In this paper we propose a In other words, the near-end and far-end speakers talking MAs with RNNs and PRNN-based computing model for AECs simultaneously or double-talk resulting modified reference problem in communication systems. Prediction of echo paths signal. can be performed by MAs-RNNs based processing. The results Non-linear processing is the removal of residual echo left about the faced problem in communication systems, where by the adaptive filter. Residual echoes are the un-modeled the use of MAs implementation coupled with RNNs might components of the echo path. Most used adaptive filters are be an improvement over a more conventional solutions, are linear and can only cancel the linear portions of the echo path. here summarized, presented and evaluated. As mentioned we Thus the nonlinear portions cannot be removed via the adaptive do not know any satisfactory solution to the echo cancellation filter and a residual echo removal follows the filter to handle problems so basically in this paper a computing model and nonlinear portions of the echo remaining. We have gained a algorithm is presented for RNNs based adaptive filtering with large experience in the development and implementation of application to AEC. This computing model can be called as echo cancellation solutions and now we use MAs with RNNS generalized MAs with RNNs-PRNN and is derived as a MAs as shown in Fig. 2. implementation. The basics of the implementation are then introduced so as the obtained results. B. Non-conventional Applications of Acoustic Echo Cancella- tion II. S OME TECHNIQUES FOR AEC The thinking of people about AEC is its application as In applications such as acoustic echo cancellation the a requirement of a conferencing speakerphones or using a impulse response of the system often reaches over 100ms wireless handset in hands-free mode. In these situations, the in length. This would require an adaptive FIR filter with loudspeaker and microphone are enclosed in the same device. over 1000 coefficients. The linear convolution and the update Therefore, the physical characteristics of the device help shape of the adaptive filter with this length creates a significant the echo path. AEC can be applied to any voice communi- computational burden for applications that require low power cation system requiring to achieve a high quality full-duplex processors. The application of the adaptive IIR filters often conversation. For example, AEC can be applied to a drive-thru fail to produce the desired results despite their reduced com- order post, home intercom systems, baby monitors, patient- plexity is because the adaptation of the IIR filter contains care intercom systems in hospitals and imaging centers, VoIP many local minima and instabilities. As the efficiency of Fast communications on laptops, videophones, and human/machine Fourier Transforms (FFTs) have improved, block processing interfaces. All of these applications present their unique set of and frequency domain adaptive filters (FDAF) were realized challenges. on low power DSPs. 110 Proc. of the 16th Workshop “From Object to Agents” (WOA15) June 17-19, Naples, Italy FDAF provide several advantages over its time domain possibly the most interesting among them is the mobile agents counterpart. Besides being able to perform the filter con- paradigm. volution by a multiplication in frequency domain, also the MAs are a special case of mobile code, i.e. processes that length of the adaptive filter are effectively decimated by the can move from one host to another and resume execution transformation. Thus, the computational complexity of the at the new host without actually restarting. There is no final adaptive algorithm is reduced. In addition FDAF can also definition of what is a MA, but attempts to classify autonomous provide an increased convergence speed. agents led someone to a general definition. The difference AEC must incorporate a sub-band adaptive filter whose between the client/server-based computing and mobile-agent- adaptation speed should be superior against conventional solu- based computing models are well described in literature. In tions As mentioned earlier, applications such as acoustic echo this latter model, instead of each sensor node sending raw data cancellation can have long echo paths, resulting in a large delay or preprocessed data to the processing center, the processing and memory requirement. This disadvantage can be overcome code is moved to the data locations through mobile agents. An by methods such as the multidelay adaptive filters. In this agents decide their own course of action, within the bounds approach the block size can be smaller than the required time of the program in the context of software agents. This is the domain adaptive filter. In this paper an MAs based algorithm meaning of the sentence: Agents are autonomous. is presented for adaptive filtering in the frequency-domain. In this section, we present the basics of the computing model based on the MAs and the following features that A. Adaptive Filtering in Echo Cancellation respond to the unique challenges posed by the sensor network. The Echo Cancellation based on Adaptive IIR Filtering is here summarized. Echo cancellation solutions are most often A. Client/server-based computing versus Mobile-agent-based based on a linear FIR adaptive filtering approach. IIR adaptive computing filters typically use a much smaller number of coefficients to Notable benefits over conventional distributed program- model a system but require additional complexity to control ming paradigms are provided by MAs. The mobile agent is a stability during adaptation. This conceptual simplicity has a special kind of “software”. Once it is dispatched, migrate from cost: even a modest approximate model of the echo path has a node to node performing data processing autonomously. The large number of filter coefficients. A very good approximation structure of a mobile agent has four attributes: identification, of the same real echo path would have much smaller number itinerary, data, and processing code. Identification uniquely of coefficients then we use RNNs and then PRNN for adaptive identifies each mobile agent. Data is the agents data buffer, filtering. In literature is reported an outline of the IIR filter- which carries a partially integrated result. Itinerary is the based echo canceller solutions. route of migration. It can be fixed or dynamically determined In the past the Kalman Filter and the Adaptive Kalman based on the current network status and the information gain. Filter for AEC was also widely used. The Kalman Filter A processing code carries out the integration whenever the was originally created in 1960 by Rudolf E. Kalman as a mobile agent arrives at a local sensor node. re-examination of the filtering and prediction problem using While in a client/server-based model, data is the migration the Bode-Shannon formalism and state-space representation unit is in the MAs based model, the migration unit is “mobile of dynamic systems. This means that the random signals to agent”. Therefore, the agent release results and terminate be worked with are represented as the output of linear systems its itinerary any time the integration accuracy satisfies the excited by white noise, and such linear systems are themselves requirement and this feature also saves both network band- described by first order difference equations. width and computation time since unnecessary node visits and agent migrations are avoided. However, for client/server- III. M OBILE -AGENT BASED C OMPUTING M ODEL based computing, there will be increased queuing delay as the number of clients increases so that result in longer processing The most commonly used computing model is called delay and more potential drops at the server side. In sensor client/server based, where individual sensors (the clients) send networks the number of nodes could be also hundreds or even raw data or preprocessed data to a processing center (the thousands. server) and data integration is carried out at the center. It’s use still resist today for distributed computing too. Some draw- The main properties of MAs are: backs are due to this computing model which might prevent it from being used in sensor networks. Firstly, client/server- 1) Scalability: The performance of the network is based computing generally requires many round trips over the not affected when the number of sensor nodes is network in order to complete one transaction. The network increased. Agent architectures that support adaptive connection needs to be alive and healthy the entire time of network load balancing could do much of a redesign the transaction, otherwise the transaction has to restart if it automatically. can at all. Secondly, some kind of super-nodes in the sensors network, served as the processing centers, have bigger storage, 2) Reliability: Mobile agents can be sent when the higher computing capabilities, and more energy. However in network connection is alive and return results some automatic and homogeneous sensor networks this is not when the connection is restablished. Therefore, the always occur being the unreliability and low bandwidth of the performance of the mobile-agent-based computing wireless link used in sensor networks. The rising demand on model is not affected much by the reliability of the processing power and the need to conserve bandwidth on large, network. slow networks claim for several new approaches appeared but 111 Proc. of the 16th Workshop “From Object to Agents” (WOA15) June 17-19, Naples, Italy Fig. 4. The selected RNN as a MA. Fig. 3. General Schematic of MAs-RNNs based computing model for N sensors node. 3) Extensibility and task adaptivity: Mobile agents can be programmed to carry different task-specific integration processes which extends the functionality of the network. 4) Energy awareness: The itinerary of the mobile agent is dynamically determined based on both the information gain and energy constraints. It is tightly integrated into the application and is energy efficient. Fig. 5. The selected PRNN as a MA for improved performance in AEC. 5) Progressive accuracy: A mobile agent always carries a partially integrated result generated by nodes it already visited. As the mobile agent migrates from methods. node to node, the accuracy of the integrated result is However during the implementation of this computing always improved assuming the agent follows the path model a few problems became apparent mainly due to the determined based on the information gain. platforms to be adopted, but the use of an MAs-RNNs ap- proach avoids the coarse approximation sometimes joined to The increasing in popularity of MAs led to the development imprecise information on the voice, echoes signals. of several programming languages specifically designed. Tele- To our problem now, i.e. as to implement RNNs in the MAs script was perhaps the first and most well known language. computer hybrid model. Initially for RNNs implementation we Java is currently the number one choice of mobile agents used RNS (Recurrent Network Simulator) which is a simulator developers. It is Javas characteristics that favour for developing for recurrent neural networks. MAs; it is inherently platform independent and a de facto RNSs features include:freely choosable connections, no standard in platform independent computing. restrictions besides memory or CPU constraints delayed links for recurrent networks, fixed values or thresholds can be speci- IV. T HE DEVELOPMENT OF THE PROPOSED MA S -RNN S fied for weights(recurrent) back-propagation, Hebb, differential AND PRNN S BASED APPROACH FOR AEC Hebb, simulated annealing and more, patterns can be specified This section documents the implementation and the simu- with bits, floats, characters, numbers, and random bit patterns lation results with the devised MAs-RNNs and PRNNs based with Hamming distances can be chosen for your user definable computing models and MAs were independent objects capable error functions, output results can be used without modification to achieve the AEC tasks. Based on users requests, the MAs as input. However we mention as is wanted the use of RNNs start their journey and move autonomously among hosts. as a guide to the agents, and then Repast (Recursive Porous Figure 3 show the general schematic for N sensors in this Agent Simulation Toolkit) is a free and open source agent scenario but overall behavior for AECwas considered in the modeling toolkit under continual development by Argonne. It paper. can be thought of as a specification for agent-based modeling services or functions. It provides an integrated set of libraries for neural networks so as genetic algorithms and other topics. A. Description Finally we favour the use of C++ for all implementedMAs in The prediction is performed at the listeners in a distribute the current our application that will be shown automatically manner using MAs-RNNs or PRNNS. The concept of the on the display. The system is started by creating a new agent entire computing model is a distributed system, then the of the Mobile Server and once created, the server shows the information is sent to the server using MAs based computing network set-up window. 112 Proc. of the 16th Workshop “From Object to Agents” (WOA15) June 17-19, Naples, Italy Fig. 6. Noisy signal on left and the signal after echo reduction on the right. V. R ESULTS (corresponding to an analysis frame-interval of 3-5.3 ms). In Fig. 2 is shown as the signal is corrupted by many types The authors have initially gained experiences by the use of noise, as the reflection in the room, or occurring in an of MAs-RNNs and these were designed as shown in Fig. 4, environment, as shown in Fig. 2. For this reason the proposed but however in the present work better MAs-PRNNs based approach should be highly reliable in presence of these noisy implementation and results are reported being several the signals. The experiments show as the average echo reduction advantages raised by their accurate implementation. An hybrid echo removing is about in the range 40 dB and 30 dB in a kind of neural network is the so called PRNN introduced by widely considered signals relative for the main speaker and Haykin and Li in 1995. It consists of a modular nested structure at several SNR ratios, for a considered signal voices database of small scale fully connected RNN and a cascaded of FIR registered at the University of Catania. By addressing the AEC adaptive filter. problem by this approach, the time is not considered as a In the PRNN configuration the single module connection is second dimension of the input space, but it is implicitly coded well described and depicted in literature, joined with relative in the structure of a recurrent network topology or modular mathematical equations. PRNN is a modular network made of RNN. a number of RNNs as its modules with each module consisting The MAs-PRNNs based echoes output prediction is so of some neurons and the selected network for the present echo obtained, but here the result is reported in Fig. 6 only for cancellation application consists of three modular layers, with a node sensor. Satisfactory performance of echo cancellation relative neuroprocessing units, as shown in Fig. 5. For both were obtained also for the other nodes. The two echo signals MAs with RNNs, or PRNNs, the number of layers, modular input, as echo remote and echo local, so as the echo output layers and neurons were defined by extensive simulation tests prediction are shown in Fig. 6. This echo output prediction in the corresponding implemented recurrent neural models. for the selected node is relative to the best result provided by The implemented MAs with RNNs, or PRNNs, have been MAs-PRNNs implementation. trained, tested using several speech voice, echoes data samples A selective management of the temporal memory is also collected from public speeches of different people and voices. possible thanks to their computational flexibility. The indi- This latter kind of network shows a better prediction capability vidual weights of the three modules in Fig. 5 are adjusted in of non linear behavior data and to model complex phenomena independent manner. Moreover PRNNs show a stronger time coming from several different physical contributes due to the memory with respect to a standard RNNs with back propaga- modular structure which provides analysis and understanding tion through time training algorithm. A large amount of data of the basic contribution to reconstruct the overall behavior. were processed in order to find a RNN and PRNN structure The signals processing is relative to signals registered at able to reproduce the behavior of a MAs soft-computing based the University of Catania coming from the microphone array. echo cancellation solutions. The MAs based echoes output For testing the MAs-PRNNs based computing algorithms a prediction in order to evaluate the echo-cancellation is so database of signals was available with a sampling rate of 8 obtained. The result were calculated for the overall system kHz. They are divided in frames of ten thousand samples 113 Proc. of the 16th Workshop “From Object to Agents” (WOA15) June 17-19, Naples, Italy keeping in account the main speaker. [9] G. Capizzi, F. Bonanno, C. Napoli, “Hybrid neural networks architectures Satisfactory performance in echo cancellation were ob- for SOC and voltage prediction of new generation batteries storage,” tained. The two echo signals input, as echo remote and echo 2011IEEE International Conference on Clean Electrical Power (ICCEP) pp. 341-344, 14-16 June 2011. local are provided as inputs to the RNN and PRNN and the [10] F. Bonanno, G. Capizzi, S. Coco, C. Napoli, A. Laudani, and G. L. echo output prediction shows a relevant reduction evaluated as Sciuto, “Optimal thicknesses determination in a multilayer structure to 60 db. This echo output prediction is relative to the best result improve the SPP efficiency for photovoltaic devices by an hybrid FEM coming from MAs-PRNNs implementation. Cascade Neural Network based approach,” 2014 IEEE International Symposium on Power Electronics, Electrical Drives, Automation and Motion (SPEEDAM) pp. 355-362, 18-20 June 2014. VI. C ONCLUSIONS [11] F. Bonanno, G. Capizzi, G. L. Sciuto, C. Napoli, G. Pappalardo, E. Tra- montana, “A novel cloud-distributed toolbox for optimal energy dispatch Several applications benefit already by the use of mobile management from renewables in IGSs by using WRNN predictors and agents technology and we now want to use a soft computing GPU parallel solutions,” 2014 IEEE International Symposium on Power Electronics, Electrical Drives, Automation and Motion (SPEEDAM) pp. technology as RNNs and PRNNs for this problem area in 1077-1084, 18-20 June 2014. communication systems. In this paper we propose a MAs with [12] C. Napoli, F. Bonanno, G. Capizzi, “An Hybrid Neuro-Wavelet Ap- RNNs-PRNNs-based computing model for AECs problem in proach for Long-Term Prediction of Solar Wind”, IAU Symposium No. communication systems. 274, Advances in Plasma Astrophysics, pp. 247-249, Giardini Naxos, Prediction of echo paths can be performed by RNNs based Italy (2010). processing. The results about the faced problem in communi- [13] C. Napoli, F. Bonanno, G. Capizzi, “Exploiting Solar Wind Time Series cation systems, where the use of MAs implementation with Correlation with Magnetospheric Response by Using an Hybrid Neuro- Wavelet Approach”, IAU Symposium No. 274, Advances in Plasma RNNs might be an improvement over a more conventional Astrophysics, pp. 250-252, Giardini Naxos, Italy (2010). solutions, are here summarized, presented and evaluated. [14] G. Capizzi, F. Bonanno, and C. Napoli, “Hybrid neural networks As mentioned we do not know any satisfactory solution architectures for soc and voltage prediction of new generation batteries to the echo cancellation problems so basically in this paper a storage, in Proc. IEEE, International Conference on Clean Electrical computing model and algorithm is presented for RNNs based Power (ICCEP), pp. 341344, 2011. adaptive filtering with application to AEC. The basics of the [15] G. Borowik, M. Woźniak, A. Fornaia, R. Giunta, C. Napoli, G. implementation are introduced so as the obtained results. Pappalardo, E. Tramontana, “A software architecture assisting workflow executions on cloud resources”, International Journal of Electronics and The experiences with the results, gained on the echo cancel- Telecommunications, vol. 61, no.1, pp. 17-23, 2015. lation problems, by the implemented MAs-PRNNs show better [16] G. Capizzi, F. Bonanno, C. Napoli, “A new approach for lead-acid performance versus MAs- RNNs computing model. However batteries modeling by local cosine” in Proc. IEEE Symposium on In- the performance of this approach can not be compared versus ternational Power Electronics Electrical Drives Automation and Motion more conventional MAs applications or distributed computing (SPEEDAM), pp.1074-1079, 2010. system in AEC problems being a novel implementation in this [17] F. Bonanno, G. Capizzi, A. Gagliano, C. Napoli, “Optimal management field but the determination if rising any advantage of using of various renewable energy sources by a new forecasting method”,in Proc. International Symposium on Power Electronics, Electrical Drives, other platform for the MAs with RNNS or PRNNs has to be Automation and Motion (SPEEDAM), pp. 934-940, 2012. investigated. [18] F. Bonanno, G. Capizzi, G. Lo Sciuto, C. Napoli, G. Pappalardo, E. Tramontana, “A cascade neural network architecture investigating surface plasmon polaritons propagation for thin metals in openmp,” Springer R EFERENCES International Publishing, pp. 22-33, 2014. [19] C. Napoli, G. Pappalardo, E. Tramontana, Z. Marszałek, D.Połap, M. [1] M. Sondhi, D.R. Morgan, J.L. Hall, “Stereophonic acoustic echo Woźniak, “Simplified firefly algorithm for 2D image key-points search”, cancellation-an overview of the fundamental problem,” IEEE Signal in Proc. IEEE Symposium on Computational Intelligence for Human-like Processing Letters, Vol. 2, no. 8, pp. 148-151 (1995). Intelligence, pp.118-125, 2014. [2] Haiquan Zhao, Xiangping Zeng, Zhengyou He, “Low-Complexity Non- linear Adaptive Filter Based on a Pipelined Bilinear Recurrent Neu- ral Network,” IEEE Transactions on Neural Networks, vol.22, no. 9, pp.1494-1507, Sept. 2011. [3] S. Haykin, “Adaptive Filter Theory,” Englewood Cliffs, Prentice-Hall, New York (2001). [4] R.H. Glitho, E. Olougouna, Pierre, Samuel, “Mobile agents and their use for information retrieval: a brief overview and an elaborate case study,” IEEE Network, vol.16, no.1, pp.34-41, Jan/Feb 2002. [5] L. Ismail and D. Hagimont “A Performance Evaluation of the Mobile Agent Paradigm” ACM SIGPLAN NOTICES, pp. 306–313, ACM Press (1999). [6] S. Haykin, “Neural Networks and Learning Machines,” Englewood Cliffs, Prentice-Hall, New York (2008). [7] F. Bonanno, G. Capizzi, C. Napoli, “Some remarks on the application of RNN and PRNN for the charge-discharge simulation of advanced Lithium-ions battery energy storage” 2012 IEEE International Sympo- sium on Power Electronics, Electrical Drives, Automation and Motion (SPEEDAM) pp. 941-945, 20-22 June 2012. [8] G. Capizzi, F. Bonanno, C. Napoli, “Recurrent neural network-based control strategy for battery energy storage in generation systems with intermittent renewable energy sources,” 2011IEEE International Confer- ence on Clean Electrical Power (ICCEP) pp. 336-340, 14-16 June 2011. 114