Raising Awareness of Artificial Intelligence for Transportation Systems Management and Operations

Chapter 2. Categories of Artificial Intelligence Technologies

Introduction

Transportation systems infrastructure owners and operators (IOO) are responsible for the planning, design, implementation, operation, and maintenance of the Nation's transportation assets. Transportation systems management and operations (TSMO) activities are frequently associated with transportation management centers (TMC).

Various TSMO programmatic areas, functions, or services within IOOs have varying degrees of sophistication in how they use analytics and data to support agency decisions. For example, pavement management programs and metropolitan planning processes within most departments of transportation (DOT) have a deep and extensive history in data-driven decisionmaking. Other programs, such as TSMO, are typically more ad-hoc in their use of data and analytics to inform key decisions. The use of performance measures, evaluating and reporting on performance, and data-driven decisionmaking processes and systems are gaining in popularity throughout IOOs.

Artificial intelligence (AI) and machine learning are elements of business intelligence (BI) strategies and technologies, which are used by enterprises for data analysis and information extraction. Traditional problems, functions, or actions that AI techniques can address include reasoning, knowledge representation, planning, learning, natural language processing (and understanding), perception, and the ability to move and manipulate objects.⁴ In each problem area, AI technologies are proving to have significant performance benefits versus other traditional mathematical modeling approaches. For example, speech recognition systems in modern digital assistants are rapidly showing improved performance over interactive voice response (IVR) technologies used in 511 systems over the past 20 years.

Various types of AI technologies can be used to help IOOs for TSMO. For example, neural networks may help agencies improve incident detection time, fuzzy logic may simplify the configuration of ramp metering systems, and natural language question-and-answering systems may expand the abilities of TSMO staff to quickly obtain information from databases of facts and figures. This document provides an overview of past research, development, and capabilities of different AI technologies. This information will also identify what may be unique for each technology, what may be a limitation, and what issues to consider when evaluating if the technology will support the needs of the agency.

In this chapter, you will become aware of:

• The distinction between "strong" and "weak" AI.
• The subcategories of "weak" AI and the basic theory of each.
• The differences between supervised and unsupervised learning.
• "Weak" AI that is relevant to various TSMO functions.

Over the last 60 years, AI research and development funding and interest has waxed and waned as the power of computers, programming languages, databases, search algorithms, and related technologies has developed. With each advancement in technology, developers are routinely optimistic about the potential for computers to replace humans in doing a wide variety of work tasks. In the early 1960s, AI funding from the U.S. and other Government agencies was substantial and the optimism that artificial intelligence would supplant many aspects of human work was substantial. When progress slowed and grandiose claims went unrealized, an "AI winter" lasted until the 1980s when computing power, memory, and database capabilities began to catch up with the dreams of AI developers.⁵

Colloquially, "artificial intelligence" typically describes the ability of a machine to mimic human actions or cognitive functions, such as problem solving or maintaining a conversation. This type of artificial intelligence is typically referred to as "strong" AI. There are no strong AI systems in existence, although progress will remain steady as digital assistants (Google Assistant, Facebook DrQA, Amazon Alexa, etc.) will continue to increase in capability to mimic what strong AI represents. Other specialized applications of AI are termed "weak" AI or machine learning applications. Machine learning applications offer the potential to supplant human work in a variety of TSMO areas, including traffic imagery analysis, incident detection, traffic control and traffic signal timing, TMC function automation, and data analysis.

General Categories of Machine Learning

Chatbots, Computational Linguistics, and Question-Answering Systems

Chatbots are AI systems that can respond to questions from a human as shown in figure 1. Until the invention of Amazon Alexa and Google Assistant, historically, chatbot interaction was through a text interface. Over the years, chatbot technology has evolved significantly, and some chatbot developers have claimed they can pass the Turing Test.^{6, 7} The Turing Test is the most famous test of strong AI competency, named after the mathematician Alan Turing. Turing's hypothesis states that if a human person could not distinguish responses in a general conversation from a computer or a person, then the artificial entity would be considered "intelligent." In practice, chatbots are only able to "pass" the Turing test in limited situations (e.g., customer service phone calls) for a limited amount of time with assumptions about the bot's command of language, breadth of education (age), and subject knowledge (e.g., history, physics),⁸ and the interrogator can only ask questions that are generally in the area in which the bot is expected to know some context.

It is generally accepted that no chatbot is yet "thinking" or represents "strong" AI. In some cases, chatbots can be remarkably good at understanding context, but in other instances the bot cannot respond to questions a four-year-old child would know. It is generally accepted that evolutionary intelligence or machine sentience is still years from being realized.⁹

International Business Machines (IBM) Watson represents the most successful question-answering system in history; it is famous for besting Jeopardy experts in 2011 and subsequent years.¹⁰ This class of AI technology is generally termed question-answering (QA) systems developed using computational linguistics. Computational linguistics has been in development since the 1970s. As illustrated in figure 2, computational linguistics seeks to map the associations of nouns, verbs, and adjectives that are generally similar.¹¹ While Watson may seem to have many properties of generalized intelligence, it is still a "weak" AI, although quite an impressive one with a massive knowledge base.

The general knowledge responses generated by Google Home, Amazon's Alexa, and Facebook's DrQA also use computational linguistics to return the closest match to your search phrase, including searching on words and concepts that mean the same thing, but are not the specific word or phrase that you provided. IBM Watson has gone on to learn how to cook and develop its own recipes, as well as finish medical school.¹² If Watson could be trained to diagnose diseases and learn the protocols of medical science, it seems possible that such an AI could be trained to solve traffic control problems. For example, pouring hundreds of thousands of microsimulation models into Watson could possibly allow it to generate the rules for how cycle, splits, offsets, and sequence relate to traffic volumes and system geometry.

Chatbots, computational linguistics, and QA systems generally represent a category of weak AI that can be useful for various TSMO functions. It is not hard to imagine systems in the near future where general questions regarding TSMO data could be presented to a QA system in natural language (assuming the necessary data elements of the question are stored in a database), such as:

• How many crashes occurred on I‑95 in Pennsylvania in August when it was raining?
• How many times was a traffic signal on Main street preempted during PM peak for more than two minutes?
• Which arterial in Montgomery County had the highest vehicle-miles traveled in June?

These questions would require several important "weak" AI technologies:

• Parsing the words of the user's question to decipher meaning.
• Generating the Structured Query Language (SQL) query to retrieve the result from the database or databases.
• Synthesizing the information and responding appropriately.

While no commercial off-the-shelf product can provide such generalized capability today, all the necessary components are currently in development or available in some form.

Expert Systems

The rise of expert systems in the mid-1980s generated renewed interest and enthusiasm for generalized AI capabilities.¹³ Expert systems represent an expert's thought processes as a set of if…then rules and logic gates (e.g., AND, OR, NOT, or NAND). Specific problem areas can be tackled when provided with enough if…then rules. A programming language called lisp was developed to represent such logical constructs, along with many capable and useful software systems that improved a wide variety of tasks. Expert systems are still in wide use; applications include medical diagnosis, toxic waste management, nuclear reactor control, and the identification of potential remedies to system failures during space missions.

The success of expert systems in specific fields led to the theory that generalized intelligence could be constructed from symbolic logic constructs and a knowledge base (i.e., a "semantic network").¹⁴ While the concept continues to be researched, the combinatorial explosion of the database size and the millions upon millions of rules needed to represent even basic elements of human knowledge was out of the scope of processing and database storage capabilities at the time.

Many of today's incident response modules of freeway management systems are expert systems. By encoding rules relating the dynamic message signs, closed-circuit television (CCTV) cameras, and other field elements to the location of an incident, TMC operators can quickly select a set of responses. Machine learning technologies may be able to generate the expert system rules more efficiently or creatively for decision-support systems.

Neural Networks and Supervised Learning

In the mid-1980s the concept of neural networks was revived by David Rumelhart in response to the rigidity of the expert system constructs of crisp logical rules.¹⁵ Loosely based on the concept of biological animal brains, neural networks "learn" to do certain tasks through training (presumably like a brain is trained as children develop) and by strengthening or weakening the connections between the network of neurons based on presented input data and the expected output result as shown in figure 3. Neural networks are commonly applied to narrowly defined pattern recognition problems, such as image recognition (is this a cat) and data segmentation (is this a cat, a dog, a fire truck, or a tuba). Most commercial AI software platforms provide a wide variety of neural network constructs for pattern recognition problems.

Machine learning algorithms are trained to solve specific problems using the general process shown in figure 4. An important element of the training process is feedback, which improves the model after implementation and evaluation of real-world results.

The goal of representing generalized human intelligence with a brain-like structure has mostly been abandoned due to computational and database limitations. However, neural networks have found significant success in solving real problems. One example is imagery analysis, which is well suited for the structure of a neural network, since many emergent features of images (such as object edges in the foreground and static scenes in the background) can be represented in the "layers" of the network.

The network of mathematical functions in a neural network is like a network of regression equations, such as y = a*x + b where a and b are coefficients fit to the data set. Neural networks have thousands of "a" and "b" parameters that fit to the data to represent vastly more complicated relationships between multidimensional x (the inputs) and y (the outputs). Neural networks and similar pattern recognition technologies are particularly well suited for problems that are noisy and multivariate, poorly represented by traditional models, and have highly nonlinear relationships between x and y.

Most neural network training through the early 2000s was supervised. This means that the correct output(s) for a given set of inputs are explicitly provided to the model. This is also known as providing "labeled" data. In imagery analysis, these labels might be that a vehicle (of any type) is in the scene or that a specific type of vehicle, such as an ambulance, is in the scene. To learn the definition of an "ambulance," many scenes with ambulances and without ambulances are provided as inputs to the model along with the correct labeling. Such pattern recognition systems are quite good at doing what you train them to do, but a common criticism is that such methods lack the ability to generalize. For example, if all images of ambulances are presented to the neural network from a side view, ambulances viewed from the front may be difficult for the neural network to correctly identify.

Many TSMO problems related to imagery analysis can be addressed by neural networks:

• Unmanned Aerial Systems (UAS) trained to identify incident details on a freeway.
• CCTV imagery analysis of crashes, debris, and pedestrians in the right-of-way.
• Counting and classifying vehicles, conducting turning movement counts, pedestrians, and cyclists.

These technologies are already on the market today and will likely continue to grow in capabilities.

Fuzzy Logic

From the late 1990s to the early 2000s, fuzzy logic was popularized after being introduced by Zadeh in the mid-1960s.¹⁷ Fuzzy logic, like neural networks was a response to the rigidity of the expert system rule bases of the 1980s. Fuzzy logic allows "if…then" rules to be probabilistic in describing linguistic variables that are modified with adjectives and adverbs as shown in figure 5. A color can be "light red" and a light can be "very bright" in human speech interactions, and most people understand such meanings intuitively; people do not describe such variables as "bright" or "red" in conversation by using lumens or red, green, and blue values. Fuzzy logic has been successful in many areas of expert systems and control systems. Neural networks and fuzzy logic are sometimes combined when the input and/or output values of the neural network are linguistic in nature. Vice versa, neural networks are often applied to tune the fuzzy rules based on past performance.

Researchers have applied fuzzy logic to a wide variety of TSMO problems and achieved some success, particularly when applying fuzzy logic to freeway ramp metering. Ramp metering algorithms before fuzzy logic required detailed and difficult to maintain models of freeway and ramp traffic conditions. By softening the rules to general descriptions of traffic conditions ("heavy traffic," "typical traffic," "light traffic"), the algorithms become much easier for humans to write and understand. This will be highlighted in chapter 4. Many AI purists posit that fuzzy logic alone does not represent AI without the learning component of neural networks or other methodologies for revision of the rules based on past performance.

Machine Learning and Solution Search

Games are a tremendously popular application of AI and machine learning methods. In 1997, IBM Deep Blue defeated chess master Gary Kasparov in one of the most famous examples of a computer built to solve a specific problem.¹⁸ The IBM Deep Blue supercomputer of 1997 is shown in figure 6. Chess (and Go, popular in Asia) is significant in the realm of machine learning paradigms because the best next move is dependent on playing out millions of potential board configurations while anticipating what an equally skilled expert opponent might do in response to each subsequent move. Deep Blue searched up to 20 rounds of potential actions in selecting the next move, mimicking generally what chess grandmasters must do in competition.¹⁹ Additional improvements to Deep Blue's strategy were made later when the AI included the recognition of board middlegame configurations and the opponent's tactical tendencies while they were playing.

Many AI purists posit that such programs for game playing should not be called AI nor even machine learning, since they "brute force" search millions of actions according to the rules of the game and do not have "intelligence," and cannot learn from their mistakes or the actions of their opponents.

Many sophisticated search algorithms have been devised over the years to improve the ability of machines to find a good (and sometimes "optimal") result much faster than iterating all possible or all feasible solutions in a given amount of search time. Such search methods include evolutionary algorithms, genetic algorithms, ant colony optimization, and a host of others.²⁰ These methods seek to resolve the problem inherent to many search techniques in that they stop searching when they reach a local maximum or minimum. This is illustrated in figure 7.²¹ Search algorithms help computer software find better solutions faster, but do not by themselves represent that the machine is learning. Much research has been done to develop search algorithms to find better solutions to common TSMO problems, such as traffic signal timings.²²

In the last 10 years, there are now many examples of game-playing programs that do learn from their actions. Google's DeepMind software, for example, can now play and master almost any Atari 2600 game from the 1970s and 1980s, and many platforming (e.g., Super Mario) and shooting (e.g., DOOM) games that have relatively simple goals, such as reaching the end of the level alive with a maximum score.²³ Delaware DOT (DelDOT), for one, is piloting the use of a "game playing" AI system to find novel solutions to integrated corridor management. This pilot program is discussed in more detail in chapter 4.

Unsupervised Learning

DeepMind and other similar systems play games by learning what actions produce better scores, such as avoiding bombs in Space Invaders or bouncing the ball behind the wall of bricks in Breakout. This type of learning is generally known as unsupervised reinforcement learning. It is unsupervised because there is no presentation of a correct output for a given set of inputs (i.e. there is no initial set of labeled scenarios that link the correct inputs to the desired output or score). It is reinforced, because each action generates a certain score; the better the action, the bigger the score. Though a process of trial and error, the software system learns what inputs (e.g. movements of the paddle in Breakout) lead to the best outputs (higher scores).

Perhaps even more impressive than playing simple Atari games, OpenAI's trained neural networks known as "five" has recently defeated and continues to best teams of professional Dota 2²⁴ e-sports players in a best of three match.²⁵ The game Dota 2 requires extensive cooperation between a team of five characters with a variety of abilities. OpenAI's "five" plays 180 years of matches against itself every day to learn strategy, tactics, and cooperation.²⁶ The achievement is considered substantial in the world of AI by Bill Gates and other industry experts as evidenced in figure 8.²⁷ The formulation of the congestion minimization problem as a "game" (i.e., TMC operator versus the network) might find some promise in coming years, since the outcomes of certain TMC operator actions cannot be known for certain.

DeepMind and OpenAI "five" are examples of the general class of machine learning methods now known as deep learning. In the case of DeepMind, the game representation is a special type of neural network, which is a commonly applied to imagery analysis since each neuron is loosely modeled after the visual cortex of a human (or animal) eye. In general, deep learning methods have multiple internal layers of feature representation. OpenAI "five" has those properties in that it "sees" the Dota 2 battlefield just like a human player would and classifies the static and dynamic objects on each image in evaluating what action(s) to take. Marketing materials for AI tools and technologies frequently use "deep learning" to loosely refer to almost any machine learning technology, but formally deep learning refers to a system that learns patterns without supervision.

Robotics and Driverless Vehicles

Robotics is the physical embodiment of machines that can substitute for humans and replicate human actions.²⁸ Robots are typically best suited for tasks that are "dull, dirty, and dangerous." Hundreds of robotic applications are commonplace in warehouses, vehicle manufacturing, explosives disposal, surgery, and agriculture, with thousands more envisioned for the future.^{29, 30} The use of UASs (though manually operated, typically) for bridge inspection, site survey, and other transportation-related activities is rapidly expanding. A prototype delivery UAS is depicted in figure 9. It is easy to envision autonomous or semi-autonomous UASs improving TSMO activities in the future. In April 2019, Google's Wing UAS received the first Federal Aviation Administration approval for package delivery.³¹

Driverless vehicles are a good example of the combination of AI techniques, robotics, computers, and sensors. The concept of driverless systems has been around since the 1930s but have only recently become viable with the current capabilities and affordability of sensors, miniaturized computers, mapping, and software systems.³² AI is used in driverless vehicles in several manners:

• Neural networks are used to detect and classify objects from light detection and ranging (LiDAR) and video inputs.
• Neural networks and similar machine learning technologies are used to fuse sensor data from multiple inputs for improving object classification (e.g., if the video and the LiDAR sensors both conclude the object is a "pedestrian," it is probably a pedestrian. If one sensor/algorithm concludes "pedestrian" and the other "bicycle," check again).
• Some AI driver models are trained to drive from watching the behavior of (good) human drivers.
• Some AI driver models are trained by unsupervised reinforcement learning in simulated three-dimensional (3D) environments.
• Anomalies in embedded maps are identified using machine learning.

Generally, driverless vehicles combine real-time LiDAR, video, and radar sensors with embedded neural network AI, high-powered computers, high-resolution digital (3D) maps, and driving algorithms (some of which also use AI models to determine what actions to take next). Driverless vehicles are still far from perfect, as evidenced by crashes with fire trucks,^{33, 34} and the recent fatalities of drivers and pedestrians.^{35, 36} Driverless vehicles will become a reality; the question is no longer "if," but "when." A driverless vehicle without a steering wheel or pedals is shown in figure 10.

The various uses and impacts of driverless vehicles on TSMO operations is commonly discussed. Speculation ranges from a utopian view of a congestion-less future and a happily ridesharing public that owns no personal conveyance, to the dystopian predictions of status quo or even more congested roadways as zero-occupancy vehicles roam the roads looking for their next fare rather than parking for 95 percent of the day as most vehicles are today.

Use of driverless vehicles and airborne UASs will likely find a variety of uses for TSMO, including:

• Crash-abatement vehicles at incident and work zone sites can be made driverless and controlled by operators with hand signals or made to follow human-driven trucks automatically.³⁷
• Incident management surveillance with airborne UASs, particularly in rural areas. Airborne imagery analysis and classification is illustrated in figure 11.^{38, 39}
• Equipment and materials delivery.
• General passenger-carrying automated vehicles relieving TSMO staff of the burden of driving enabling more productivity en route to job sites.
• Automated classification of ground objects from 3D LiDAR point clouds using neural networks.⁴⁰

Summary

AI and machine learning technologies will likely enable IOOs to increase the sophisticated use of analytics and data to support TSMO activities. Chatbots and QA systems may enable new ways to obtain insights in data. Neural networks will analyze imagery from a variety of sources for incident detection, incident management, and traffic data collection. Fuzzy logic is already used by a variety of DOTs for ramp metering and they may find additional applications by simplifying if…then rule bases for decision-support systems. Unsupervised AI systems may learn new ways to control traffic and coordinate integrated corridor management actions across a variety of control and advisory technologies. Driverless vehicles and airborne and ground-based drones will likely improve TSMO staff safety and productivity. Additional applications are likely to arise as agencies gain experience with AI tools and technologies.

No generalized human-like intelligence (or "strong" AI) is available today or envisioned to be available anytime soon. Machine learning technologies are trained by supervised or unsupervised methods. Neural networks are the most popular form of machine learning methods, particularly for imagery analysis. Significant investments in driverless vehicles and UASs will likely see near-term applications that will affect a wide variety of TSMO functions.

The next chapter discusses some commercial AI computing platforms. As commercial companies continue to quickly add to their system capabilities, feature sets will continue to evolve, systems will continue to become easier to use, and the cost of ownership or use will decrease. This view of commercial capabilities should be considered just a snapshot of early 2019. The use of product names and specific tools does not constitute an endorsement of these technologies.

You may need the Adobe® Reader® to view the PDFs on this page.