Over the last two decades, computers have become omnipresent in daily life. Their increased power has enabled machine-learning techniques, reliable natural language processing, strong human-computer interfaces, and other capabilities that now allow computers to handle analytical tasks that were solely the domain of humans until quite recently. For example, today’s navigational software helps drivers choose the shortest route to their destinations, using a combination of GPS, maps, and current traffic conditions, and it can interpret and respond to verbal commands. The step of allowing computers to actually drive a vehicle on public roads may not be far off. Similarly, computerized decision making facilitates power delivery to our homes, keeps planes in the air, and alerts our physicians about our health risks.
Along with this increasing computer power, networking technologies now make it possible for people and computers to access enormous stocks of information worldwide. Search technologies, recommender systems, database technologies, and other tools are some basic capabilities for discovering relevant information. Our ability to automate workflows and exploit distributed computing allows systems to marshal more processing power and data than most individuals or enterprises actually own and control.
This state of affairs has generated an environment that is ripe for a re-thinking of human-computer collaboration in the context of complex decision making. However, the vast amount of information that can be brought to bear does not guarantee better decisions or a more straightforward decision-making process. To build that capability, scientists, engineers, and technologists need to address a broad range of challenges as described in this report. We need a stronger foundation of knowledge in order to efficiently and reliably share tasks between humans and computers.
For example, computers are more capable than humans in finding and “digesting” huge amounts of data, and that attribute is exploited in systems such as modern electric grids, which are adjusted rapidly in response to changes in loading, and in “fly-by-wire” aviation. But the decision-making capabilities of machines are limited, in part, because their models of factors influencing decisions are limited; for instance, their models can only partially represent such human cognitive abilities as seeing the whole picture in context, including special circumstances (Hoffman et al., 2002); and being able to incorporate the implications of unexpected events and situations. Machines cannot consider the full range of decision options in part because their sensors and databases cannot provide inputs for every contingency, but also because science lacks sufficiently complete understanding to be able to foresee and include all of the variables that people may consider important. Automation can be fragile, subject to failure through mechanical problems, programming limitations, or
contradictory, erroneous, or otherwise inappropriate data. Sensors measure whatever they are capable of measuring, which is seldom the actual information of most relevance to the decision maker. Human beings can be inconsistent at decision making, sometimes displaying excellence, and sometimes being prone to errors and biases. Recognition of this inconsistency has led humans to find ways to improve decision making by formalizing the process, such as through analytical methods and improved understanding of how humans and teams address decision making.
The committee believes it has been shown that technology can genuinely improve complex decision making by humans in many situations. Major computational advances over the past 2 decades have resulted in forms of computational support (data collection, data analysis, computing power, visualization, etc.) that have been readily incorporated into human decision making processes.
The specific statement of task given to the committee reads as follows:
Conduct an analytical assessment of global research efforts in several technologies that enable humans, machines and computer systems to collaboratively digest, analyze and act on vast amounts of unstructured data in dynamic environments. This analytical assessment will include findings on
(1) key research goals in several enabling sub-disciplines that support human-machine decision making,
(2) main impediments to achieving technological breakthroughs,
(3) key systems-integration challenges, and
(4) the scope and character of international approaches to these research areas.
The sub-disciplines to be studied include, but may not be limited to: brain–computer interface, machine learning, natural language dialogue systems, sensing and perception, software agents, and cognitive and social science issues. The committee will produce a report with findings but no recommendations.
The study committee’s expertise spanned the disciplines named in the Statement of Task, but its deliberations quickly revealed the difficulty in trying to bound this challenge. The following are examples of additional dimensions that are relevant to the topic under study, but which the committee decided not to cover in order to maintain focus:
- Non-technical factors that influence the success of decision making and of teams, such as emotion, social context, culture, relationships, organizational structures, authority systems, and so forth. Many such factors can lead to the failure of team decision making and are not avoided simply by improving the interactions between humans, computers, and networks. For example, great strides are being made in the development of coordination tools for distributed teams, which addresses one aspect of this. Another example is how culture affects attitudes and the politics of a region, and hence decision making. Related to that is the question of how cultural differences affect reactions to technologies, such as the degree to which new technologies are trusted and accepted. There is a good deal of emerging research on this topic.
These topics are mentioned at the end of Chapter 2’s section “Overview of Decision Making,” but much more could be said.
- Another aspect that could be explored more deeply is the use of feedback from decision-making teams to improve the structure and operation of decision-making processes. This is analogous to the way Facebook makes decisions about its features, interface, and interactions with people based on how its subscribers interact with one another. Big data offers the potential to improve our understanding of human networks and interactions, thus altering and enhancing the way complex decision making is managed. While the report discusses a number of ways in which big data affects (now, or potentially) human-machine coordination in complex decision making, this is an emerging area, and much more could be said.
- More generally, the report does not attempt to characterize the state of the practice of exploiting big data for decision making. A complete examination would encompass aspects such as the major approaches to learning from big data (e.g., supervised vs. semi-supervised vs. unsupervised learning) and assessing the progress and promise of various approaches (e.g., neural networks, support vector machines, Bayes graphical models). Instead of delving into these topics, the report cites and quotes from a 2013 National Academies report on the subject, Frontiers in Massive Data Analysis.
- The committee did not incorporate a specific discussion of human-robot collaboration and interaction because the issues involved center more around the consequences of autonomy than around the aggregation of information and coordination of team decision making. In their well-established applications, such as in factory lines, the decisions made by robotic systems are fairly prescribed and not particularly analogous to the types of interest to this study (see next chapter). However, consequences such as safety hazards1 are persistent due to the robots’ inability to sense and mitigate such hazards. Emerging applications of autonomous machines, such as self-driving vehicles and military drones, rely on more advanced decision-making capabilities2, and thus additional concerns arise due to the possibility that actions may be taken (and properly executed) based on imperfect decision-making, perhaps with tragic consequences. These challenges cannot be properly addressed in this report.
- The committee did not delve deeply into ways to aid the cognitive work of sensemaking and computational models of attention. Some key references for these topics are introduced in the section on Human Cognition and Memory in Chapter 3 and in the section on Neuroscience in Chapter 5.
This report explores ways that people and computer systems can collaborate so that complex decisions involving large amounts of data and tight time constraints are better
1See, e.g., John Markoff and Claire Cain Miller, As robotics advances, worries of killer robots rise, The New York Times, June 17, 2014.
2See the National Research Council 2014 report, Autonomy Research for Civil Aviation: Toward a New Era of Flight. Washington, DC: The National Academies Press.
made. The immensity of the topic prompted the committee to give priority to outlining the context in which human-machine collaboration for decision making can profitably be discussed. This raises many research issues concerning team decision making among humans, and then similar challenges are identified when the “team” is extended to include humans and machines. In addressing these topics, the report discusses the nature of the research, achievements, and systems-integration challenges of several enabling technologies that underlie human-machine collaboration for decision making.
A 2012 workshop at the National Academies explored the topic of intelligent human-machine collaboration. It is instructive so see the range of ways in which participants at that workshop answered the question, “What Is Intelligent Human-Machine Collaboration?” The following sample of responses was quoted in the published summary of that workshop:
“… machines and humans combining each other’s strengths and filling-in for their weaknesses and empowering each other’s capabilities;
“… joint and coordinated action by people and computationally based systems, in which each have some stake in the outcome or performance of the mission;
“… humans AND machines jointly perform tasks that they would not be able to perform on their own;
“… integration of AI into machines;
“… humans and machines are able to mutually adapt their behavior, intentions, and communications;
“… cooperation that mimics interactions between two humans;
“… naturalness of the observed human-machine interaction;
“… neither human nor machine treats the other as a disturbance to be minimized.
“… machines being partners, and not a tool, for humans;
“… technology that amplifies and extends human abilities to know, perceive, and collaborate;
“… better overall performance of the mission, independently of how it was achieved;
“… shared responsibility, authority, goals.”3
Overall, one can see the breadth of this topic and the absence of precise definitions and boundaries.
This report begins by examining the kinds of decisions that motivated the study. They are largely characterized by the availability of large amounts of information of varied types, which introduces a certain type of complexity. (Other types of decisions can be very complex for other reasons, such as the need to balance a range of perspectives and/or appropriately consider serious or sensitive consequences. The study that led to this report devoted less attention to such drivers of complexity.) Having laid out that context, the report discusses major human, computer, and network elements of team decision making. It then surveys the research frontiers that provide the basis for human-machine collaborative
3National Research Council. Chiang, E. N., and. Wrightson, P. S., rapporteurs. 2012. Intelligent Human-Machine Collaboration: Summary of a Workshop, The National Academies Press: Washington, D.C. Quoted material from p. 2 (ellipses in original).
decision making, with an emphasis on building an integrated view, and offers thoughts on future directions for such research.
More specifically, Chapter Two provides the context for the question that informs this study: How might humans and computers team up to turn data into reliable (and when necessary, speedy) decisions? Here the committee looks at the three basic components of this question: the essence of decision making; the vast amount of data that have become available as the basis for complex decision making; and the nature of collaboration that is possible between humans and machines in the process of making complex decisions.
Chapters Three and Four examine, respectively, the human elements and the machine and network elements of team decision making. Chapter Three addresses several aspects of human decision making that can benefit from computer-aided systems, such as cognition, errors in judgment, and task allocation. Chapter Four focuses on the teaming of humans and computers to make decisions, and ends with a discussion on how metric classes might contribute to advancing human-machine decision making. Chapter Five looks at research areas that underlie human-machine collaboration for decision making: sensing, software agent systems, neuroscience, and human computation.
Chapter Six contains observations about where this research may be headed. Findings related to research opportunities are included in that chapter, while other findings appear throughout the report where appropriate.
In Appendix B, the report recounts committee visits to research organizations in both Singapore and Germany, but it does not assess the quality of research.; The goal of these overseas site visits was to address item (3) in the study charge, concerning “the scope and character of international approaches to these research areas.”
Throughout this document, the terms and words “computational system,” “computer,” “machine,” “information system,” and “automation” are used interchangeably to make the document more readable. Given that people are experts at interpreting context, each of these words or their derivatives are also used in ways that are not interchangeable in a few places.