Over the past two decades, questions have surfaced about the effectiveness of contribution of intelligent systems to decision makers in a variety of settings. With the proliferation of such systems in operational settings, such as aerospace, medical, manufacturing, and transportation systems, increased attention to evaluation of such systems, and to resulting software safety, became warranted (Grabowski & Sanborn, 2001). Multiple studies evaluating such systems and creating evaluating criteria for future studies have been produced, and the majority of them find that the old, more simplistic approach of one problem-one solution in such computer system is becoming obsolete. Instead, the systems are evolving into self-learning and sometime self-replicating mechanisms, with a capacity to seek solutions to non-standard problems while encountering them only in theory rather than in reality (Fazlollahi & Vahidov, 2001; Baydar & Saitou, 2001). In other words, the machines are learning to think like humans: solving problems via flexible, rather than rigid, approach; observing various operational errors and how the applied solutions affect them; and acquiring the ability to use theoretical thinking, or even imagination.
In operational sense, acquiring such abilities is essential for such computer systems, since their main purpose is to replace humans in certain areas. For example, the embedded intelligent real-time systems (EIRTS), which are knowledge-based systems deployed in larger host systems with real-time response requirements, are used in the maritime industry for navigational purposes (Grabowski & Sanborn, 2001). The task to be supported by the host system and the EIRTS is ship’s piloting, a cognitively complex task comprised of three activities – tracking, maneuvering and collision avoidance, and the practice of good seamanship (Grabowski & Sanborn, 2001). In order to perform these operations, the EIRTS utilize three types of information – local knowledge, ship-handling knowledge, and transit-specific knowledge (Grabowski & Sanborn, 2001). All these operations are traditionally performed by humans, namely the ship’s crew. What the computer systems need to learn, therefore, is how to perform these operations in the same manner the humans do, meaning that not only all the external and internal factors need to be taken into account, but also that one course of action must be chosen out of the number of possible ones.
In order to achieve this, the systems must be supplied with enough sensor receptors so as to be able to obtain maximum available information that may influence the movement of the ship through water, such as condition of the sea, traffic, speed, water currents, and so on. In the case discussed by Grabowski and Sanborn (2001), such sensors are available to the EIRTS in the form of system sensors, or depth sounders, radars, navigation sensors, radio direction finders, the Global Positioning System, the ExxBridge integrated bridge system, and the Shipboard Piloting Expert System (SPES) (Grabowski & Sanborn, 2001).
Having external sensors by themselves, however, is not enough to perform the necessary functions. What is needed for proper operations is a memory, the knowledge of the layout of the land and sea, to which the variables obtained through sensory observation can be applied. The human equivalent would be the map of a certain area that is memorized from both personal use and communication with other travelers and which is then used for travel without the need to consult the actual printed map. In the described case, such backbone for the system is the SEANET, a high-speed, real-time network connected to the network modules by network interface units (NIU) (Grabowski & Sanborn, 2001).
The combination of sensory input and the existing navigational data allows the EIRTS to make actual conclusions about any situation and to devise a plan of reaction by reasoning which one of the available action would be most favorable to the vessel. The functional tasks of the EIRTS system include collecting data from SEANET; performing situation assessment and plausible action generation activities to detect and predict risks of collision or grounding; and developing recommended courses of action in the form of voyage plan alterations (Grabowski & Sanborn, 2001). These tasks mirror the bridge watch team’s tasks with respect to track keeping, maneuvering and avoidance, and the practice of good seamanship; so essentially, the computer is performing all the same operations humans do, and subsequently must think and make decisions in the same manner humans do (Grabowski & Sanborn, 2001).
Even at this stage, however, the system setup is not complete. In order to operate effectively, a human being must be able to communicate with his or her colleagues, both to share information and recommendations and to receive additional input adjusted to reflect the changing conditions. This means the computer system must be equipped with means to communicate its findings and conclusions. In this case, EIRTS display their alerts, alarms, and recommendations as graphical overlays on the ExxBridge electronic chart – for example, as a flashing icon indicating collision danger with a target – as text output of recommendations and explanations, and as auditory signals associated with alerts and alarms (Grabowski & Sanborn, 2001).
Potential advantage of computer-performed operations over human-performed operations may have been considered to be the machine’s apparent mental incorruptibility due to the lack of emotional stress. For simpler, individual computer systems, this may be true, but in the discussed case, the EIRTS are embedded systems, which means they are part of a bigger host and depend on it not only in ways of sharing information, but on its performance as well. They are constrained by their host’s parameters, requirements, and performance (Grabowski & Sanborn, 2001), just like the human brain is constrained by the chemical processes inside the human body. Therefore, the performance of the EIRTS may diminish in quality if the host system slows or degrades. But just like human brain can evaluate the condition of the body in which it resides and convey its conclusions to others, so are EIRTS capable of notifying humans who work with it on the bridge about any glitches in the operations of its host (Grabowski & Sanborn, 2001).
In addition to internal damage, many real-time systems may be subject to external physical damage from their environments, such as accidents in an industrial automation application, or hostile attack in case of a defense application. This is why real-time systems, such as EIRTS, are often evaluated in terms of fault tolerance, recovery, redundancy, patterns of faults, failures, and errors, and software performance under time pressure.
Since EIRTS are intelligent systems, there must be developed a way to evaluate their performance that would satisfy their complex nature. In this case, verifications of intelligent systems are conducted at three levels. One is a task level, where the system is evaluated on what it does, why, and what requirements are satisfied by the system’s operation (Grabowski & Sanborn, 2001). Another one is the representation level, where evaluation of logical organization of coding structures used to represent knowledge and reasoning in the system is made (Grabowski & Sanborn, 2001). The third level is that of implementation – evaluating the algorithms, representations, and knowledge used to build the system (Grabowski & Sanborn, 2001). The human equivalent of such verification would be creating proficiency exams in various subjects and conducting them together with taking into account the examinee’s intellectual background.
Intelligent systems tend to become so much more like humans, their evaluation proceeds to the level of technology used to create a system and includes analysis of how the system impacts individual or group processes such as decision making or workload and communications. Just like a more gifted or a more reasonable person would make more sensible decisions that someone with less ability, so are quality decision-making processes, supported by the superior technology, more likely to result in high-quality decisions (Grabowski & Sanborn, 2001).
Perhaps the most important attribute of human reasoning is the ability to ‘think on the fly,’ or to address each problem or task in a loose, flexible manner, with as few rigid parameters as possible, utilizing all new available information. The ability of a computer system to revise its conclusions as new data becomes available is also verified in this case at the representation level (Grabowski & Sanborn, 2001).
The unreliable nature of a rigid system of evaluation and testing applies equally to human and computer operations. For example, an intelligent system’s performance can be verified by humans at the implementation level by a series of test cases, or by checking system’s knowledge bases for internal correctness (Grabowski & Sanborn, 2001). However, each such test is very rigid in terms of determining the level of system’s performance, and two separate tests can produce vastly differing results because of the differences in their underlying concepts of correctness. Some of these aspects are the absence of certain phenomena, correctness with respect to test cases, or even logical correctness, and they are often specific to a particular knowledge representation scheme or inference strategy (Grabowski & Sanborn, 2001). In some studies, a knowledge base is determined to be ‘correct’ if certain functional and structural properties, such as conflict or disagreement with case data, are absent from the knowledge base (Grabowski & Sanborn, 2001). Because potentially there could be such a vast difference in evaluations of the same system if done by different methods, this type of evaluation does not appear reliable or practicable.
Similarly, when it comes to determining the errors in the automated assembly lines, programmable logic control (PLC) codes, which comprise 90 percent of all the control coding effort aimed at error recovery, are programmed only on the basis of ‘expected’ error scenarios (Baydar & Saitou, 2001). They are deficient in dealing with ‘unexpected’ scenarios, leaving the recovery process to manual labor work and ultimately defeating the purpose of having a testing program in place altogether. Similarly, the simplicity of the ‘what if’ simulation defeats the entire purpose of conducting a simulation, even though running theoretical scenarios in order to determine how to handle problems that had not yet come up is in itself an advancement for the intelligent computer systems (Fazlollahi & Vahidov, 2001).
This is where another aspect of reasoning, which was previously attributed exclusively to humans, comes into play, which is experience. The EIRTS performance is routinely tested on response time and variability criteria with respect to combinations of input conditions in order to establish an intelligent system’s workload. While input conditions vary, the system compares them against those already stored there and utilizes the existent patterns of reasoning if the conditions match favorably (Grabowski & Sanborn, 2001).
For the automated assembly lines, the solution to the problem of limited error-recovery capacity of the PLC is a special program that is composed of several commands and directions for the industrial robot. This program can be downloaded to the robot controller to perform the recovery task. These commands in the recovery program are chosen from an available set of commands, so while the recovery robot does not learn anything new, it still draws on the downloaded ‘experience’ that is broader than a single-purpose error scenarios of the PLC (Baydar & Saitou, 2001).
Perhaps the most telling proof that the latest-generation computer systems are attempting to copy the human data-processing and decision-making approach is the latest implemented technology on the market, which is genetic programming. It effectively creates systems with self-replicating and self-adjusting testing parameters, which allows them not only to implement new data faster, but also to improve automatically as they experience the data on which they are trained (Baydar & Saitou, 2001).
One of the main advantages of genetic programming is that all recovery plans can be generated directly in the language in which operating system is written. In other words, there is no need to translate the code into the language understandable to humans and then for humans to input the data back into the system using the language they recognize and then having it translated into a code recognizable by the machine (Baydar & Saitou, 2001). For this discussion, however, the element of bigger importance is that operating system runs tests and then adjusts its error recovery protocol according to the procedures that took place as a result of rectifying that error, without the outside help.
The nature of genetically programmed systems is very similar to how the human genome works. In genetic algorithms, design variables are coded onto fixed-length or variable-length strings that are analogous to chromosomes in biological systems (Baydar & Saitou, 2001). Sixteen strings are composed of characters that are analogous to human genes. Each string represents a solution point in the search space. An objective function is defined within the problem, and the genetic algorithm tries to maximize the fitness of a solution point based on a fitness function related with the objective (Baydar & Saitou, 2001).
What makes such method of programming so similar to the human genes is the way the algorithms change themselves in order to adjust to error variations. In genetic algorithms, two basic operators are applied to the selected pairs. First, the ‘crossover,’ makes the strings of each two members cut and recombined from a random point, thus producing two new members. Secondly, a randomly selected place on a randomly selected string can have its value altered, hereby creating a ‘mutation’ (Baydar & Saitou, 2001).
Similarities with the way human or any other organic genes change themselves are uncanny. Both changes serve precisely the same purposes, both in the world of organics and cybernetics. Mutation in the algorithms has the advantage of introducing some diversity into the search for errors, as well as ways to rectify them (Baydar & Saitou, 2001). Organic mutations represent the way living things on this planet adjust to changing conditions, be it weather or any other. True, computer mutations are immeasurably more frequent then organic ones, but ultimately only those that fit the need of the current situation get implemented. On the other hand, crossover uses the properties of the current population to combine and produce better results. (Baydar & Saitou, 2001). Farmers and scientists all over the world use similar method to improve the quality of livestock and crops, for example.
Some aspects of genetic development of cybernetic systems can be used as confirmation of reasons for the Earth’s biodiversity. In the genetically programmed systems, the fitness value of both the best and the worst recovery programs is increasing as the evolution of the algorithms takes place (Baydar & Saitou, 2001). A computer system may use the worse recovery programs to rectify a simpler error, and, if there are variables, to improve that program. It will not be better than the best program the system uses, but if called upon to recover more and more errors, under constantly varying conditions, the program would improve to the point where it could be considered the best the system has. When applied to the organic world, such development successfully explains why some organisms developed one set of limbs or appendages better than the other, or why their bodies are shaped in some specific way. In all the cases, with no exception, every change in appearance of an organism is caused by the body attempting to solve the problems related to the organism’s environment on the most basic, genetic level. So, if one part of the body is challenged to ‘recover errors’ more often than another, under constantly varying conditions, it would end up being developed to the highest degree, or, in other words, would be considered the best recovery program the organism has created. Meanwhile, the part of the organism that would be challenged to the smallest degree would either receive less genetic adjustment or be adjusted in such a way as to minimize its interference with the other, more useful parts. However, even while being inferior to the better organic recovery programs, the underdeveloped one still would be able to recover errors for which it has been created successfully. In addition, should the new challenges arise in that area, the organism would adjust the program through a mutation or a crossover.
Finally, modern intelligent computer systems possess something that no other organism except humans does, which is imagination. The cybernetic imagination comes in a form of three-dimensional simulations that the systems run in order to improve their capacity for error recovery without actually being involved in error recovery. Such simulations exists in every article under analysis here, and its main purpose is for the computer system to learn how to take the initiative in performing decision-related tasks by increasing its ‘memory,’ or database of possible scenarios, potential errors in them, and recovery methods (Fazlollahi & Vahidov, 2001).
The importance of running such simulations is clear, especially in an economic sense. It is considerably cheaper to spend electricity and programming hours to create simulated trouble situation than have the real ones arise and incur the cost of slowed production, or failed trip and late delivery, for example, while the computer systems learns the proper recovery programs from its own mistakes. The results of such an approach could be catastrophic.
From this analysis, it is clear that using the methods utilized by human for their decision-making and problem solving is the best alternative for computer systems designers. After all, nature had a much longer period of time in which to experiment with its organic versions of computers than humans had to improve on their own cybernetic creations.
Baydar, C.M. & Saitou, K. (2001). Automated generation of robust error recovery logic in assembly systems using genetic programming. Journal of Manufacturing Systems, 20(1), 55-68.
Fazlollahi, B. & Vahidov, R. (2001). A method for generation of alternatives by decision support systems. Journal of Management Information Systems, 18(2), 229-250.
Grabowski, M. & Sanborn, S.D. (2001). Evaluation of embedded intelligent real-time systems. Decision Sciences, 32(1), 95-123.