Intelligent Methods for Embedded Systems
1Kumar Keshamoni 2V. Rajitha Rani 3A. Ravichandra
Senior Assistant Professor, Dept. of ECE Associate Professor, Dept. of ECE Senior Assistant Professor, Dept. of ECE
Aurora's Scientific Technological & Research Academy Aurora's Scientific Technological & Research Academy Aurora's Scientific Technological & Research Academy
Chandrayanagutta, Telangana, India Chandrayanagutta, Telangana, India Chandrayanagutta, Telangana, India
Kumar.keshamoni@gmail.com rajita_voggu@yahoo.co.in ravichandra.aleti@gmail.com
Abstract ' This paper discusses auspicious strategies for the implementation of intelligent solutions for embedded systems. Associate embedded system could be a automatic data processing system designed to perform an infatuated or slender vary of functions with a minimal user intervention. associate intelligent system could be a system that's able to react fittingly to dynamical things while not user input. Main challenges for intelligent solutions in embedded systems return from reliability and time period necessities and from constraints on value, size, and power consumption. Possible intelligent methods for embedded systems are biologically inspired, such as neural networks and genetic algorithms. Multi-agent systems are also prospective for an application for non-time critical services of embedded systems. Another field is soft computing which allows a sophisticated modeling of imprecise (sensory) data. Finally, since embedded systems often provide critical services, there is need for intelligent validation techniques that assist the developer in evaluating if the system is fit for its purpose.
1. INTRODUCTION
Designing embedded systems is quite different from desktop programming. Desktop pro-grammers are able to use standard environments with almost unlimited (virtual) memory as well as a good monitoring and debugging interface. In contrast, an embedded pro-grammer has to renounce such comfort. He or she has to cope with many different micro-controllers, some of them providing only several kBytes of program memory and a few bytes of working memory. Instead of a comfortable environment with graphical screen display and printers, developers have to use LEDs, the display of an oscilloscope, or a serial data stream for debugging. Embedded systems should run on sparse resources, thus should require low power, little program and working memory, be small in size, and often should guarantee real-time behavior or be resistant against failures. Many embedded ap-plications have been implemented as 'dumb' programs, consisting only of a few lines of code. However, the embedded market calls for more extensive and smarter applications, for example, an electronic braking system of a car should be able to provide its service in various extreme situations, e. g., the breakdown of a braking element of one wheel. Therefore, a new generation of intelligent embedded systems is necessary. Most algo-rithms for intelligent embedded systems already exist in the computer science community,
However, due to the fundamental difference between desktop and embedded computing, many approaches need to be reviewed in order to determine if they are applicable for embedded systems. It is the objective of this paper to provide an overview on intelligent methods that are prospective for miscellaneous tasks of embedded applications. The remaining parts of the paper are structured as follows: Section 2 provides the necessary definitions for intelligent systems and discusses some misconceptions around that term. Section 3 lists a number of possible benefits for intelligent solutions in embedded systems. Section 4 explains, why biologically inspired systems are a good template for intelligent embedded systems. Section 5 explains the concept of multi agent systems and refers to some applications of multi agent systems in the embedded domain. Section 6 motivates and proposes the con-cept of soft computing for embedded systems. Section 7 discusses the methods of model checking and fault injection for system validation. The paper is concluded in Section 8.
2. DEFINITION OF INTELLIGENT SYSTEMS
'Intelligence' refers to the overall effectiveness of an individual's mental processes, par-ticularly his or her comprehension, learning/recall, and reasoning capacities. When in-telligence is seen as the capability to solve (new) problems, it is possible to identify 'in-telligent' solutions in engineering. It should be made clear, that 'intelligent behavior' in this context does not refer to the ability to solve puzzles or to be conscious of ourselves being. These goals of a strong artificial intelligence are far above the capabilities of our current machines and computers. Intelligence in engineering means systems that are able to react appropriately to changing situations without input from a human operator. In other words, an intelligent algorithm is one that is able to solve problems that stem from changing situations. This definition, of course, encompasses a wide range of engineer-ing applications and many different methods and algorithms. The research on intelligent systems is motivated by the high versatility of such systems, which makes it possible to reuse many algorithms successfully in different applications. There are some misconceptions around intelligent systems. For example, a system that does not have intelligent behavior on its own is probably nevertheless the product of an intelligent mind, a fact that might be used by the advertising department to call the product itself intelligent. Moreover, an intelligent system might not be the best solution for a given application case. For example, a racing bicycle is specifically designed for good roads, while the human legs adapt well to various terrains like grassland, water, mountains, etc. Therefore, the human locomotor system represents an intelligent system while a racing bicycle is none. Nevertheless, it's in all probability advantageous to settle on the bike in some cases.1 At last, associate degree intelligent system should not be essentially complicated. Leslie Smith [1] describes the instance of a automaton that navigates towards a lightweight supply. Primitive biological systems tend to resolve these forms of tasks in an exceedingly easy, however effective, way.
3. MOTIVATION FOR INTELLIGENT EMBEDDED SYSTEMS
Naturally, the word 'intelligent' (as well as good, wise, clever) transports a awfully positive that means, however, it's obvious that it ought to be analyzed why exploitation intelligent solutions for embedded systems is advantageous. The following potential reasons for using associate degree intelligent resolution are often identified:
Dependability: Applications for harsh environments like method management applications require an answer that adapts to dynamic things like performance loss or break- down of a part. For such applications, intelligent solutions change swish
degradation or self-stability properties.
Efficiency: An intelligent solution might be able to increase efficiency of the given re-sources.
Autonomy: An intelligent solution might be able to perform the same task as a traditional system without or with reduced requirement for human supervision or interaction.
Easy Modeling: An intelligent generic self-organizing solution liberates the system de-signer from modeling and implementation issues. This reduces the chance of human error and reduces cost and time in the design phase.
Maintenance costs: An intelligent system might require less frequent service iterations since it is able to run for long durations without human interaction.
Insufficient alternatives: Sometimes there is no traditional approach to solve a given problem satisfyingly, which forces the application of an intelligent solution. For example in data analysis, the application of neural networks solves the problem of nonlinear correlations, which is not supported by traditional approaches [2].
Note, however, that the modelling issue gives also reason for a counter argument against intelligent solutions. While a traditionally designed system usually must be understood by the designer, this can be different for particular intelligent solutions. For example, while neurons are well understandable, their application in a neural network leads to an emer-gent system which cannot be fully described by a simple model. If such a system should be used in a critical application, problems for the dependability analysis and certifications arise.
As a positive, almost all above described potential advantages support also cost reduction. Such cost savings can appear at the design time of a system or during maintaining the system.
4. BIOLOGICALLY INSPIRED EMBEDDED SYSTEMS
Nature has shown to be a great inventor of intelligent solutions in embedded systems. The reasons for that fact are given by two major requirements that are put on biological systems: First of all, most biological systems have to work autonomously. Usually, there is no one to help a creature in recovering from a breakdown. Therefore, animals, as well as plants, had to develop strong methods of self-healing and automatic recovery. Of course, there is positive interaction between individuals: Mammals protect their new-generation, many animals live or hunt in groups, and there is the concept of symbiosis that involves even multiple different species. However, once concerning any of those teams of inter depending people, a system of sturdy inner affiliation that's able to facilitate itself will be created out. The second demand that biological systems have in common with embedded engineering systems is that almost all parts area unit used to perform a fanatical or slender vary of functions. Such biologically embedded systems area unit continually optimized to similar goals as embedded laptop systems: creation effort, maintenance effort, size, weight, power consumption.
4.1 Neural Networks
The most frequent example for biologically inspired computing is that of neural networks (NNs). A NN consists of interconnected neurons, each with a set of input and output connections. In principle, a neuron contains a simple add-and-compare mechanism that sums up the input signals and generates an output signal (i. e., the neuron 'fires') if a particular threshold has been exceeded. While the concept of such a neuron cell is very simple, a whole NN shows emergent properties such as learning and reasoning (for ex-ample, think of the abilities of the human brain, a NN with about 1010 neurons). NNs are extreme versatile. According to the theorem of Hecht-Nilsson [3], any given function can be expressed by a three-layer NN with an appropriate number of neurons. An impressive feature of a NN is the ability to learn, which enables such systems to adapt to changing conditions [4]. NNs support supervised and unsupervised learning. In supervised learning, back-propagation NNs are used. During a training phase, the parameters of the NN are adapted until the system performs the desired function. The trained system is then able to perform the programmed function with a high robustness against errors. An example for such an application in the embedded domain is given in [5], where an artificial NN is used to filter out errors from infrared distance sensors. Unsupervised learning algorithms try and extract common sets of options from the input file [4]. associate example for associate unattended learning artificial NN is Kohonen's selforganizing map [6]. unattended learning algorithms area unit used for automatic classification, modeling, and knowledge compression systems.
Drawbacks of NNs area unit its black-box processing structure and, in some cases, a slow convergence speed. Thus, the info process mechanism of a NN can't be programmed, understood, or verified in terms of rules.
4.2 Genetic Algorithms
A genetic algorithmic program (GA) may be a derivative-free and random improvement methodology that builds on concepts from the survival and therefore the organic process [7]. It's some reasonably search algorithmic program that's advantageous if the given search house is just too massive to be searched by thorough search algorithms and too unstructured to be ready to use simple search algorithms. Moreover, a GA desires solely a minimum on info concerning the matter to be solved and is therefore simply applied. Basically, a GA needs an initial population of 'genes', an algorithm that allows to cross-mix these genes, and a fitness function that produces a comparable value on the quality of an actual solution. After recombination and mutation of genes the GA uses the fitness function to select the best genes for the new population. By making multiple iterations, the GA approaches an solution that is equal or better than the start value.
An example for the application of a GA, which is relevant for embedded systems, is given by Atanassov in [8]. The work focuses on the search for an input set with the maximum (worst case) execution time (WCET) of a given program on a given execution environment. Knowledge of the maximum execution time of a program is essential for several hard real-time architectures, such as the Time-Triggered Architecture [9]. The search for the WCET input set is non-trivial due to the usually large input space and the in homogeneity of the search space due to code dependencies and side effects. Atanassov's approach estimates the WCET of a program by measuring the execution time with sev-eral input sets and then modifying the input sets towards the maximum execution time. He uses a GA for finding input sets with large WCETs, whereas the fitness function is given by the execution time with the respective input set. The evaluation of this approach shows that the genetic algorithm is able to find a rather good solution in relative short time (Atanassov's experiments ran for several days on an embedded C167 processor), however, it reveals also an inherent drawback of GA – since the search is not exhaustive, the algorithm can get stuck in local extrema, thus fails to find a global optimum of the fitness function. In other words, GAs are usually very fast in finding a good solution, but in general they will not find the best solution.
4.3 Neuro-Fuzzy Systems
Fuzzy Logic forms a bridge between digital rules (for example 'if measured flow is larger than fifty l s , then open valve') and inaccurate info (for example 'flow is between forty eight and fifty two l s '). The illation technique of formal logic is comparable to the human brain. formal logic supports the implementation of management algorithms for inaccurate sensors that perform higher than ancient management ways. associate degree thoroughgoing introduction into formal logic with a spotlight on management ways will be found in [10].
However, Fuzzy Logic has the drawback of lacking an effective learning mechanism 'auto-tuning a classical Fuzzy system is difficult [7]. The combination of Fuzzy systems with neural networks overcomes some problems of NNs and Fuzzy Logic, by providing an adapting system with a rule-based model. Such neuro-fuzzy Systems employ learning algorithms of a NN to determine the parameters of a Fuzzy inference system. Unlike NNs, a neuro-fuzzy system is always interpretable in terms of fuzzy if-then rules, thus giving insight into the model.
5. MULTI-AGENT SYSTEMS
Wooldridge and Jennings define a multi-agent system (MAS) as hardware or (more usually) software-based computer system that provides the following properties [11]:
Autonomy: Agents operate without the direct intervention of humans or others, and have some kind of control over their actions and internal state.
Social ability: Agents interact with other agents (and possibly humans) via some kind of agent-communication language.
Reactivity: Agents perceive their environment, (which may be the physical world, a user via a graphical user interface, a collection of other agents, the internet, or perhaps all of these combined), and respond in a timely fashion to changes that occur in it. Pro-activeness: Agents do not simply act in response to their environment, they are able to exhibit goal-directed behavior by taking the initiative. The idea of a multi-agent system (MAS) is to interconnect several widely independent agents, thus enabling this ensemble to function beyond the capabilities of a single agent of the set-up [12]. In general, MASs may enhance speed (due to parallelism), reliability (due to redundancy), efficiency, and flexibility. A frequent paradigm for an MAS is an automated travel agency that uses the internet to engage other agents, to perform flight reservations, etc. Recent research has also shown the applicability of MAS to the em-bedded systems domain. An example is the PABADIS project [13], which employs an MAS in a fieldbus system for factory automation [14]. There are, however, some critical problems for MAS:
Communication: Agents must agree on a common transport protocol and a common communication language in order to interact properly with each other [12].For example, an MAS that spans different embedded systems probably will have to deal with differing data representations and semantics.
Integration of existing applications: Especially in the embedded systems domain many applications come in the form of legacy applications, i. e., systems that have been designed according to their own rules and conventions [15]. Since the participation of such systems was not in their designers' mind, there can be remarkable effort in implementing an appropriate interface for MAS.
Real-Time Capabilities: Due to the loosely coupling of the single agents, agents of an MAS are typically asynchronous and, therefore, prone to race conditions, temporal unpredictability, and, in the worst case, to deadlock situations.
Supervision: It is often difficult to monitor the behavior of every single agent with respect to real time. Monitoring tasks can change the behavior of an MAS due to the so-called probe-effect [16].
Despite these problems, MAS are apt to implement intelligent functions for embedded systems. Due to the problems in temporal predictability, applications for MAS lie mainly in non time-critical applications, such as, for example, configuration tasks.
6. SOFT COMPUTING
The world of computers are digital, and at a very low level, only 0 or 1 exist (or true and false). The properties in the real world, on the other hand, are often different from that black-and-white thinking, which often is the reason for problems in embedded systems, where a system views its components either as correct (according to its specified service) or incorrect (outside the service specification). So, a component that provides its service only a little bit offside its specification might be perceived as correct or incorrect by a correct system, which causes a divergence in the correct system states. To overcome such real life complexities such as imprecision the computer paradigm of soft computing is used. Soft computing is associate degree overall term for a coalition of methodologies such as formal logic, neuro-computing, organic process computing, probabilistic computing, chaotic computing and machine learning [17]. Neural Networks, Genetic Algorithms, and Neuro-Fuzzy systems, that are already mentioned during this paper within the context of biologically-inspired computing will be also seen as a kind of soft computing. Since many soft-computing approaches are also biologically inspired, there is a great intersection of these two disciplines. Apart from the digitalization problem, sensor measurements get an additional dimension by regarding their accuracy in the value and time domain, and, since sensors can fail, the reliability of a particular measurement.
Embedded systems often contain sensors that form the borderline between the digital computer world and its analog environment. Instead of reducing a sensor measurement to its value, it is often advantageous to add some additional information about the probability of the given value. Buede and Waltz discuss the benefits of such soft sensors and probabilistic sensor fusion in [18]. An architecture for sensor fusion with probabilistic measurements can be found in [19], where each sensor measurement is attributed with a confidence value that indicates the accuracy of the measurement.
7. MODEL CHECKING
Usually, intelligent solutions result in complicated systems, wherever it's usually tough to prove the correctness of the given answer. for instance, a neural network consists of the many neural nodes observed as neurons, where each neuron acts deterministically per its input. However, once the neural network is taken into account a full, it's hard to prove a particular behavior. Thus, there is a want for intelligent ways in which assess if a given system is appropriate its purpose. A accomplishable collateral technique is model checking, a way for collateral finite state synchronic systems. The system-under-test is modelled by a group of state variables and additionally the accomplishable transitions that will occur between the particular states. a group of properties distinguishes the meant (allowed) states from the unintended states. A model checking tool will then automatically search the accessible state house therefore on verify that exclusively meant states are going to be reached. the most disadvantage of this approach is that the state explosion drawback that happens once verifying systems with high correspondence or massive information domains. Therefore, the state explo-sion drawback has been target of analysis that LED to some remedies for the matter, to name a few: symbolic model checking [20], abstraction [21], symmetry [22], and induc-tion [23]. In distinction to ancient approaches, like simulation, testing, and generalization, model checking is kind of fast in detecting refined bugs for several applications [24]. As a bonus, the model checking technique forever presents a disproof if a such property does not hold for a given vogue. once setting up the model, the checking formula is completely automatic and thus desires no user direction for its application, that creates it associate intelligent tool. as a result of recent analysis, model checking is further and extra used for embedded systems [25].
8. SUMMARY AND CONCLUSION
All presented methods, summarized can be potentially used for the conception, design and utilization of intelligent systems for particular embedded applications. There is no all-round solution for an intelligent solution – it will be necessary to evaluate from case to case if an approach is propitious or not. Furthermore, we should keep in mind that it is possible to use most methods in a synergistic way, which may lead to the most favorable solution in some cases. Some existing disadvantages, such as the resource requirements on memory and com-putation for implementing a neural network can be overcome by applying an appropriate hardware-software co-design which is a common approach for many embedded designs. However, a major problem for many intelligent solutions is that they come in the form of a complex system, which cannot be easily evaluated analytically – it is often difficult to check whether a system is fit for its purpose or not. This problem calls for intelligent validation methods like model checking.
References
[1]. L. S. Smith. Biologically-Inspired Systems. Department of Computing Science and arithmetic, University of Stirling, Scotland, Scotland, UK, 1999. Lecture notes on Biologically galvanized Com-puting.
[2]. E. C. Malthouse. Limitations of nonlinear PCA as performed with generic neural networks. IEEE Transactions on Neural Networks, 9(1):165-173, January 1998.
[3]. N. Nilsson. Learning Machines. Morgan Kaufmann, San Mateo, 1965.
[4]. D. D. Lee and H. S. Seung. Learning in intelligent embedded systems. In Proceedings of the Work-look on Embedded Systems, Cambridge, MA, USA, March 1999.
[5]. H. Kraut. Signalverarbeitung mittels eines Neuronalen Netzwerkes f ''r einen sensible device. Bache-lor's thesis, Technische Universit''t Wien, Institut f ''r Technische Informatik, Vienna, Austria, 2003.
[6]. T. Kohonen. Self-Organizing Maps. Springer, Heidelberg, 1995.
[7]. X. Z. Gao and S. J. Ovaska. Soft computing ways in motor fault dianosis. Applied Soft Computing,1:73-81, 2001.
[8]. Atanassov, S. Haberl, and P. Puschner. Heuristic worst-case execution time analysis. In Proceed-ings of the tenth European Workshop on Dependable Computing, pages 109-114. Austrian laptop Society (OCG), May 1999.
[9]. C. Scheidler, G. Heiner, R. Sasse, E. Fuchs, H. Kopetz, and C. Temple. Time-Triggered design (TTA). Advances in info Technologies: The Business Challenge, IOS Press, 1997.
[10]. W. Pedrycz. Fuzzy management and Fuzzy Systems. John Wiley & Sons, 1993.
[11]. M. Wooldridge and N. R. Jennings. Intelligent agents: Theory and apply. data Engineering Review, 10(2):115-152, 1995.
[12]. H. S. Nwana and D. T. Ndumu. A perspective on code agents analysis. data Engineering Review, 14(2):1-18, 1999.
[13]. The PABADIS syndicate. Revolutionising Plant Automation The PABADIS Approach, 2002. Avail-able as study at http://www.pabadis.org.
[14]. Y. K. Penya, S. Mahlknecht, and P. R ''ssler. light-weight agent platform for superior field-bus nodes. In Proceedings of the Work-in-Progress Session of the fourth IEEE International Workshop on manufacturing plant Communication Systems (WFCS'02), page 71, V''ster''s, Sweden, August 2002.
[15]. C. Jones, M.-O. Killijian, H. Kopetz, E. Marsden, N. Moffat, D. Powell, B. Randell, A. Ro-manovsky, R. Stroud, and V. Issarny. Final version of the DSoS abstract model. DSoS Project (IST-1999-11585) Deliverable CSDA1, October 2002. out there as analysis Report 54/2002 at http:/ World Wide Web.vmars.tuwien.ac.at.
[16]. C. E. McDowell and D. P. Helmbold. Debugging coincidental programs. ACM Computing Surveys,21(4):593-622, December 1989.
[17]. L. A. Zadeh. Applied soft computing – prolusion. Applied Soft Computing, 1:1-2, 2001.
[18]. D. M. Buede and E. L. Waltz. advantages of soppy sensors and probabilistic fusion. Signal and information Processing of little Targets, Proceedings of the SPIE, 1096:309-320, 1989.
[19]. W. Elmenreich. device Fusion in Time-Triggered Systems. PhD thesis, Technische Universit''t Wien, Institut f ''r Technische Informatik, Vienna, Austria, 2002.
[20]. K. L. McMillan. Symbolic Model Checking: associate degree Approach to the State Explosion downside. PhD thesis, faculty of technology, Carnegie philanthropist University, 1992.
[21]. E. M. Clarke, Jr., O. Grumberg, and D. E. Long. Model checking and abstraction. ACM Transactions on Programming Languages and Systems, 16(5):1512-1542, 1994.
[22]. E. A. Emerson and A. P. Sistla. Symmetry and model checking. Formal ways in System style,9:105-131, 1995.
[23]. K. S. Namjoshi. amelioratory the State house Explosion downside. PhD thesis, University of Lone-Star Stateat Austin, 1998.
[24]. E. M. Clarke, Jr., O. Grumberg, and D. A. Peled. Model Checking. MIT Press, 1999.
[25]. S. Gnesi. Model checking of embedded systems. ERCIM News, (52), March 2003.
[26]. M. T. Jensen. Improve strength and suppleness of timing and total flow-time job outlets exploitation robustness measures. Applied Soft Computing, 1:35-52, 2001.
[27]. A. Ibrahim and B. Nath. planning best neuro-fuzzy systems for intelligent management. In Proceed- ings of the sixth International Conference on management, Automation, Robotics, and Vision, December2000.