Advantages and limitations of Artificial intelligence
Advantages and Limitations of Artificial intelligence
By Md iftekharul islam
BICE 18.ID NO.18511113
Artificial Intelligence (AI) is the branch of computer sciences that emphasizes the development of intelligence machines, thinking and working like humans. For example, speech recognition, problem-solving, learning and planning.
Artificial intelligence (AI) applications are utilized to simulate human intelligence for eithersolving a problem or making a decision. AI provides the advantages of permanency,reliability, and cost-effectiveness while also addressing uncertainty and speed in either solving aproblem or reaching a decision. AI has been applied in such diverse realms as engineering,economics, linguistics, law, manufacturing and medicine, and for a variety of modeling,prediction, and decision support and control applications (1). One of the most promisingapplications of AI has been its rigorous use in the Internet such as in search engines (2).Although the efficacies of AI are significant, as with any application they are limited in bothcapability and functionality. These limitations will be presented later in this article. Beforedescribing AI’s limitations, this article will briefly survey some of AI’s advantages.In an organization in which human intelligence is tied to a particular person or a group ofpeople, AI applications can provide permanency that prevents the knowledge from being lostwhen the individual or the group members retire or are no longer available to the organization.The life of the knowledge encapsulated in an AI framework could be as long as the relevance ofthe problems and decision scenarios remain unchanged. AI also enables the development of alearning capability which can be utilized to further prolong the life and relevance of theapplication. Learning from real-world success and failure is an enabling feature of AI toolsknown as “reinforcement learning” and is advantageous in that it increases the reliability of thetools with their increased use in applications (2).The broad application of any tool only occurs when its reliability has been established,and AI has already proven to be quite reliable in many different applications because of itsability to simulate human intelligence in a reasoning process. Like many automations, AIsupports cost minimization as it enables reduction on the need of personnel time. An agency canreduce significant staff time by adopting appropriate AI applications in the decision-makingprocess, thus reducing operational costs. As decisions must often be made under obvious uncertainties (i.e., with incomplete anduncertain knowledge), AI methods are suitable when a direct mathematical relationship cannotbe established between cause and effect. AI models capture the uncertainty between real-life cause and effect scenarios by incorporating available knowledge with probabilities andprobability inference computations (3). AI methods are also capable of dealing with
bothqualitative as well as quantitative data, a feature that most strictly analytical methods lack.Depending upon the computational time in terms of algorithmic complexity and processor capacity, AI tools can facilitate faster decision making by automating the decision-making process. Through data gathering and screening, processing, and decision making, AI cansupport faster solutions to complex problems.
In transportation, numerous research and applications have demonstrated many of the above advantages of AI in general, and significant research has provided evidence of theadvantages of specific AI tools. Some examples of AI technology currently in use includeconverting traffic sensors into intelligent agents that can automatically detect and report trafficaccidents or predict traffic conditions (4). More recently, researchers have found AI to be morereliable in assessing and predicting traffic conditions, based on microscopic traffic data collectedfrom vehicles on their path, as envisioned in the vehicle–infrastructure integration or theconnected vehicle program, compared to many other existing algorithms. Utilizing microscopictraffic data, transportation security is another realm in which AI can be of significant use (5).Here, AI tools can be applied to the identification of security breaches, and in the developmentand management of automated response and control plans. The well-known advantages and efficacies of AI make them particularly useful in thedevelopment and management of transportation systems. Specifically, in intelligenttransportation systems, real-time sensing, detection, response, and control are of paramountimportance, and AI can be utilized effectively in all of these applications. Indeed, a distributedtraffic sensor and control network is perhaps the future of real-time traffic management andcontrol. Here, distributed sensor networks, comprised of different level of intelligent sensor networks, automatically detect and respond to incidents and control roadway network as needed.Such an intelligent sensor network comprised of AI tools can support the development of the next generation traffic management system. With the many advantages of AI tools, we expecttheir wider adoption in different areas of transportation. Nevertheless, one should not forget that, like any other tool, AI methods have their limitations. One major criticism of many AI paradigms (e.g., neural networks), which was previously alluded to in the article by van Zuylen, is that they are often regarded as black boxesthat merely attempt to map a relationship between output and input variables based on a trainingdata set. This also immediately raises some concerns regarding the ability of the tool togeneralize to situations that were not well represented in the data set. One solution that has beenproposed to address the black box problem is the combination or integration of multiple AI paradigms into a hybrid solution (e.g., combining neural networks and fuzzy sets into neuro-fuzzy systems) or coupling AI paradigms with more traditional solution techniques.Another limitation of AI-based search methods, such as genetic algorithms and ant colony optimization, is that they are never guaranteed to reach the “optimal” solution. Also when using AI-based search methods to solve a problem, it is often hard to gain true insight into the problem and the nature of the solution, as is possible for example when using mathematical programming methods. The inability to quickly do sensitivity analyses is an important example of this limitation. The counter argument for the inability to guarantee optimality is that for hard optimization problems that defy solution using traditional optimization and mathematical programming methods, a “solution” is still better than “no solution” at all. Moreover, there is significant empirical evidence to suggest that AI-based search methods do yield “good” solutions
in most cases. For gaining insight into the problem, the model may have to be rerun multiple times to assess the sensitivity of the solution to the various assumptions and parameters of the problem, which may be somewhat demanding from a computational resources or runtime standpoint. A third limitation associated with the use of AI methods to solve a given problem stems from the fact that, for several AI methods, there is currently little guidance on how to decide upon the best values to use for a given method’s tuning parameters. Considering neural networks, for example, the analyst needs to make some very important design decisions including the topology of the network to use, the number of hidden layers, the number of neurons to use in the hidden layer, the type of trigger functions to use in the neurons, among others, before applying this method to a given problem. In current practice, an analyst often has to go through a tedious trial-and-error procedure to select appropriate values for these parameters. The same is true for genetic algorithms, where the analyst needs to decide upon the population size, the number of generations, and the algorithm’s other control parameters such as the probabilities for applying the mutation and crossover operators.With respect to automation in general, and AI is no exception, there is also another limitation or a challenge related to the issue of potential liability. For example, if in the future, AI methods are used to build partial or full autonomous vehicles, for example, and if an autonomous vehicle is involved in a crash, who should be held liable in such a case? While this limitation is not technical in nature, it is a serious issue that needs to be addressed.
Finally, if we are to specifically focus on AI applications in transportation, one has tonote that there is currently still a lot of skepticism among transportation practitioners regarding the ability of AI to help solve some of the problems they face. One reason behind such skepticism may in fact be attributed to the very name of the field itself—“artificial intelligence”—and the rather unrealistic goals for the field adopted by early AI researchers. As this committee has pointed out in Transportation Research Circular E-C113: Artificial Intelligence in Transportation: Information for Application, our purpose, as transportation researchers and professionals, is not to develop a general-purpose problem solver, but rather to address real transportation problems that have defied solution using classical solution methods.The following part of this circular will therefore briefly survey the wide range of transportation problems and applications areas where AI techniques have been applied. The part will also identify additional opportunities for applying AI methods. Specifically, five broad application areas where AI has been applied will be reviewed. These are (a) traffic operations; (b) travel demand modeling; (c) transportation safety and security; (d) public transportation; and(c) infrastructure design and construction.