Human-Machine Inference Networks For Smart Decision Making: Opportunities and Challenges
HHUMAN-MACHINE INFERENCE NETWORKS FOR SMART DECISION MAKING:OPPORTUNITIES AND CHALLENGES
Aditya Vempaty † Bhavya Kailkhura ‡ Pramod K. Varshney ∗† IBM T. J. Watson Research Center, Yorktown Heights, NY ‡ Lawrence Livermore National Laboratories, Livermore, CA ∗ Department of Electrical Engineering and Computer Science, Syracuse University, Syracuse, NY
ABSTRACT
The emerging paradigm of Human-Machine InferenceNetworks (HuMaINs) combines complementary cogni-tive strengths of humans and machines in an intelligentmanner to tackle various inference tasks and achieveshigher performance than either humans or machines bythemselves. While inference performance optimizationtechniques for human-only or sensor-only networks arequite mature, HuMaINs require novel signal processingand machine learning solutions. In this paper, we presentan overview of the HuMaINs architecture with a focuson three main issues that include architecture design, in-ference algorithms including security/privacy challenges,and application areas/use cases.
Index Terms — human-in-the-loop systems, behav-ioral signal processing, self-driving cars, health care in-formatics, intelligent tutoring systems
1. INTRODUCTION
In traditional economics, cognitive psychology, and ar-tificial intelligence (AI) literature, the problem-solvingor inference process is described in terms of searchinga problem space, which consists of various states of theproblem, starting with the initial state and ending at thegoal state which one would like to reach [1]. Each pathfrom the initial state represents a possible strategy whichcan be used. These paths could either lead to the desiredgoal state or to other non-goal states. The paths from theinitial state that lead to the goal state are called the solu-tion paths. There could be multiple such paths betweenthe initial and the goal state which are all solutions to theproblem. In other words, there are multiple ways to solvea given problem. The problem-solving process is to iden-tify the optimal (under a given constraint) solution path
This work was performed under the auspices of the U.S. De-partment of Energy by Lawrence Livermore National Laboratory un-der Contract DE-AC52-07NA27344 (LLNL-CONF-742883) and sup-ported in part by the AFOSR DDDAS program under grant FA9550-17-1-0313. among the multiple solution paths emanating from theinitial state and reaching the goal state.The first step for such a search is to determine the setof available strategies, i.e., the strategy space. The sec-ond step is to evaluate the strategies to determine the beststrategy as the solution. In traditional economic theory,a rational decision maker is assumed to have the knowl-edge of the set of possible alternatives , has the capabil-ity to evaluate the consequences of each alternative, andhas a utility function which he/she tries to maximize todetermine the optimal strategy [2]. However, it is widelyaccepted that humans are not rational but are bounded ra-tional agents. Under the bounded rationality framework[2, 3], decision makers are cognitively limited and havelimited time, limited information, and limited resources.The set of alternatives is not completely known a priori nor are the decision makers perfectly aware of the conse-quences of choosing a particular alternative. Therefore,the decision maker might not always determine the beststrategy for solving the problem.On the other hand, machines are rational in the sensethat they have stronger/larger memory for storing alterna-tives and have the computational capability to more ac-curately evaluate the consequences of a particular alter-native. Therefore, a machine can aid a human in fast andaccurate problem-solving. This leads us to a frameworkfor human-machine collaboration for problem-solving.In this paper, we discuss this collaboration frameworkfor inference by defining the Human-Machine InferenceNetworks (HuMaINs) and discuss the research challengesassociated with developing such a framework. The threebasic threads of research in this area are defined.
2. HUMAIN FRAMEWORK
Fig. 1 presents a typical Human-Machine Inference Net-work (HuMaIN). A typical HuMaIN consists of a so- The terms strategy and alternative are used interchangeably. The terms sensor and machine are used interchangeably. a r X i v : . [ c s . H C ] J a n ial network where humans exchange subjective opinionsamong themselves, and a machine network where ma-chines exchange objective measurements amongst them.Moreover, due to the interaction between social and ma-chine networks, the behavioral characteristics of humansdetermine algorithms adopted by machines and these al-gorithms in turn affect the behavior of humans. There-fore, an intelligent collaboration of humans and machinescan deliver improved results, by exploiting the strengthsof humans and machines. Social Network Interaction Sensor Network Interaction
Human-Machine Inference Network
Subjective Opinions Objective MeasurementsAlgorithm DesignBehavioral Change
Fig. 1 . Notional HuMaINs architectureThere are three major directions of research that fallunder the HuMaIN paradigm: architecture, algo-rithms, and applications.
3. ARCHITECTURE
Several control architectures involve the interaction of anautonomous system with one or more human agents. Ex-amples of such architecture include fly-by-wire aircraftcontrol systems (interacting with a pilot), automobileswith driver assistance systems (interacting with a driver),and medical devices (interacting with a doctor, nurse, orpatient) [4]. The success of such architecture dependsnot only on the autonomous system, but also on the ac-tions of the human agents. The goal is to develop a hu-man decision-making framework that quantifies the hu-man representation in the decision-making task under un-certainty, and also develop an estimator for the model pa-rameters. The framework should also provide a commonontology for humans and machines to share relevant in-formation about the task. By estimating the parameters,a machine can access this representation and potentiallyimprove its performance. In control systems terminol-ogy, the model and associated estimator should form aplant-observer pair for human decision making that canbe used for system design [5]. Incorporating these ideasinto the feedback control framework will require new re-sults and theory to provide performance guarantees.We classify the architectures into three categories: architectures where humans directly control the au- tonomous system, architectures where the autonomoussystem monitors humans and takes actions if required,and a combination of and . In order to achievethe goal of HuMaINs, it is critical to build an architec-ture that lends itself to a blend of human and machinedecision making. In [6], it is stated that to create a state-of-the-art operator environment for modern automationsystems, continued technology development is needed inthree major areas: decision support tools; ergonomicsand visualization technologies; and ease-of-use of com-plex systems. Research focusing on building such sys-tems falls under the research paradigm of Human-in-the-loop cyber-physical system (HiLCPS) [7]. As Schirneret al. [7] state, designing and implementing a HiLCPSposes challenges that requires multi-disciplinary researchto solve these challenges. Research in the areas of con-trol systems, human-computer interface (HCI), and sys-tems design, together will drive the design of a HuMaINarchitecture.
4. ALGORITHMS
The key research area for HuMaINs is the developmentof new algorithms that deal with the human-behavioraldata. This falls under the paradigm of an emerging re-search area called behavioral signal processing [8]. Be-havioral Signal Processing (BSP) deals with human be-havioral signals. It is defined as processing of human ac-tion and behavior data for meaningful analysis to ensuretimely decision making and intervention (action) by col-laborative integration of human expertise with automatedprocessing. The goal is to support and not supplant hu-mans [8]. The core elements include quantitative under-standing of human behavior and mathematical modelingof interaction dynamics. Narayan and Georgiou describethe elements of BSP by using speech and spoken lan-guage communication for measuring and modeling hu-man behavior [8].
Observe human behavior Build behavioral models(Re)Design human-machine systems
Fig. 2 . General approach for the design and analysis ofHuMaINs.There are two specific research directions while de-veloping BSP algorithms for HuMaINs:. Develop mathematical models of human decisionmaking using statistical modeling techniques, inclose collaboration with cognitive psychologists,and2. Design robust fusion algorithms that handle unreli-able data from the agents as modeled by the abovedeveloped models.These problems have both theoretical and implementa-tion challenges. Both these research problems are furtherdiscussed in some detail below.
The first step towards developing efficient systems con-taining humans and machines is to develop appropriatemodels that characterize their behavior. While statisticalmodels exist that characterize the machine observations,researchers have not extensively investigated the mod-eling of decisions and subjective confidences on multi-hypothesis tasks, or on tasks in which human decisionmakers can provide imprecise (i.e., vague) decisions. Bothof these task types, however, are important in the manyapplications of HuMaINs. In the preliminary work [9], acomparative study between people and machines for thetask of decision fusion has been performed. It was ob-served that the behavior between people and machinesis different since the optimal fusion rule is a determinis-tic one while people typically use non-deterministic ruleswhich depend on various factors. Based on these obser-vations, a hierarchical Bayesian model was developed toaddress the observed behavior of humans. This modelcaptured the differences observed in people at individ-ual level, crowd level, and population level. Movingforward, for individual human decision-making models,tools from bounded rationality framework [3] and ratio-nal inattention theory [10] can be used in building a the-ory. Experiments with human subjects can be designed tomodel the cognitive mechanisms which govern the gen-eration of decisions and decision confidences as they per-tain to the formulation of precise and imprecise deci-sions. One can also build models that consider the ef-fect of stress, anxiety, and fatigue in the cognitive mech-anisms of human decision making, decision confidenceassessment, and response time (similar to [11, 12]).
The next step after deriving probabilistic models of hu-man decision-making, is to develop efficient fusion al-gorithms for collaborative decision making. The goalwould be to seek optimal or near-optimal fusion ruleswhich incorporate the informational nature of both hu-mans and machines. Due to the large volume of data in some practical applications, it is also of interest toanalyze the effects that a large number of agents (hu-mans/machines) and a high rate of incoming data have onthe performance of the fusion rules. However, the highlyparameterized nature of these human models might deemtheir implementation impractical. Also, the presence ofunreliable components in the system might result in poorfusion performance. Data from existing studies in thecognitive psychology literature along with models result-ing from the work in Sec. 4.1 can be used in the analysisof these operators. For cases in which the implemen-tation of the optimal rule is not feasible, one must in-vestigate the use of adaptive fusion rules that attempt tolearn the parameters of the optimal fusion rule online.Also, for the design of simple and robust algorithms,ideas from coding theory can be used similar to the re-liable crowdsourcing results such as in [13].For the development of future systems consisting ofhumans and machines, the methodology described aboveneeds to be implemented. First, statistical models of hu-mans should be developed, which are then used to opti-mize the machines in the system. Due to the presence ofpotential unreliable agents, one has to also take into con-sideration the robustness of the systems while developingsuch large-scale systems. For example, [13–15] demon-strated the utility of statistical learning techniques andtools from coding theory to achieve reliable performancefrom unreliable agents.
5. APPLICATIONS
Another major driver for the development of large-scaleHuMaINs is the application areas. Each application areathat deals with human-machine collaboration has its ownspecific nuances that drive the architecture and algorith-mic solutions. In this paper, we discuss four extremelyimportant and timely application areas: education, au-tonomous vehicles, health-care, and science. We discusstheir associated research problems in the context of Hu-MaINs.
Human-in-the-loop system can have a significant impactin education domain. The research field of IntelligentTutoring Systems (ITS) is attempting to design computersystems that can provide immediate and customized in-struction or feedback to learners, with intervention froma human teacher. They are enabled to serve as comple-menting a human teacher and ensure personalized andadaptive learning at scale to every learner. While ITS re-search has been active for several decades [16,17], recentadvancements in AI and big data research has enabled in-reasingly more human-like interactions with computersgiving rise to interactive, engaging, and immersive tutor-ing systems. A typical ITS consists of four basic compo-nents [18, 19]: Domain model, Student model, Tutoringmodel, and User interface model. The Domain Modelcontains the skills, concepts, rules, and/or problem-solvingstrategies of the domain to be learned. The Student/LearnerModel is an overlay on the domain model and it modelsthe student’s cognitive and affective understanding of thedomain and their evolution during the learning process.The tutor model represents the tutoring strategies and ac-tions that are dependent on the domain model and thespecific learner. The user interface component integratesthe other three to ensure interaction with the user andlearning advances as planned.With respect to the HuMaIN paradigm discussed inSec. 2, the domain model represents the task or goal of aHuMaIN, research on the user/learner model representsthe human aspect of HuMaINs and modeling human be-havior, the tutoring model represents the machine aspectof HuMaINs and designing of robust inference algorithms,and the user interface model represents the architecturalresearch of designing HuMaINs.
Detection, localization, control, and path planning are es-sential components of autonomous vehicle design [20,21]. These tasks focus on sensing and interacting withthe physical world through sensors and actuators. Al-though, autonomous vehicles can be a game changer, thereare still many obstacles holding back their deploymentin practice. Autonomous nature of these systems makethem quite vulnerable to cyber-attacks. A solution is toemploy human-in-the-loop systems (semi-autonomous driv-ing) for safe and intelligent autonomous vehicle oper-ation. Such systems would require joint environment-driver state sensing, inference, and shared control andnew metrics to characterize safety. The measures of sys-tem safety should take into account human performancein response to unexpected hazardous events, and humandecision making during vehicle malfunctions caused bycyber-attacks. Furthermore, allowing communication amongmultiple self-driving cars can enable collective intelli-gence in such systems, however, would require the de-sign of robust communication protocols.
Automated inference using machine learning (ML) forhealthcare holds enormous potential to increase quality,efficacy and efficiency of treatment and care [22]. Au-tomatic approaches greatly benefit from big data withmany training samples. Several tasks in medical domain have high dimensional complex data, where the inclu-sion of a human is impossible and ML shows impressiveresults. On the other hand, for certain tasks one is con-fronted with a small number of data sets or rare events,where ML-approaches suffer from insufficient trainingsamples. Furthermore, in healthcare, decisions made bymachines can have serious consequences and necessitatethe incorporation of human experts’ domain knowledge.There is also a growing trend of litigation requiring theneed to bring human in the loop. This makes doctor-in-the-loop systems to be a perfect candidate for healthcare.Designing such systems would require devising ML ap-proaches that can interact with human agents (doctors)and can optimize their learning behavior through theseinteractions. Furthermore, unlike current black-box likeML approaches, we need interpretable ML models forhealthcare so that these systems can become transparentto earn experts’ trust and be adopted in their workflow.
Scientific research spans problems and challenges rang-ing from screening of novel materials with desired per-formance in material science, optimizing the analysis ofthe Higgs boson in high energy physics, tracking of ex-treme weather phenomena in climate science. Currently,the role of machines in accelerating science has been lim-ited to solving a well-defined task where the data andtechniques are given to them by the scientists. This limitsour ability to tackle problems where not only the com-plexity of the data but the questions and the tasks it-self challenge our human capabilities to make discover-ies [23]. HuMaINs can play an important role in sci-entific research, and become crucial as more interdisci-plinary science questions are tackled. This would requireadvancing machine learning techniques to do indepen-dent inquiry, proactive learning, and deliberative reason-ing in the presence of hypotheses, domain knowledge,and insights provided by the scientists.
6. CONCLUSION
In this paper, we presented an overview of human ma-chine inference networks. Specific attention was paid tothree main issues: 1) architecture design, 2) inference al-gorithms, and 3) application areas. A holistic researchinitiative across these issues is needed to empower thisnew field of HuMaIN research. Also, moving forward,the social aspect of HuMaINs, with multiple human andmachine components interacting, such as in IoT systemsis a direction for future research. . REFERENCES [1] J. R. Anderson,
Cognitive Psychology and Its Im-plications . New York: Worth Publishers, 2010.[2] J. G. March and H. A. Simon,
Organizations . NewYork: Wiley, 1958.[3] H. A. Simon,
Models of Bounded Rationality: Em-pirically Grounded Economic Reason . Cam-bridge: MIT Press, 1982.[4] W. Li, D. Sadigh, S. S. Sastry, and S. A. Seshia,“Synthesis for human-in-the-loop control systems,”in tacas2014 , Apr. 2014, pp. 470–484.[5] P. B. Reverdy, “Human-inspired algorithms forsearch: A framework for human-machine multi-armed bandit problems,” Ph.D. dissertation, Prince-ton University, 2014.[6] J. Pretlove and C. Skourup, “Human in the loop,”
ABB Review , vol. 1, pp. 6–10, 2007.[7] G. Schirner, D. Erdogmus, K. Chowdhury, andT. Padir, “The future of human-in-the-loop cyber-physical systems,”
IEEE Computer , vol. 46, no. 1,pp. 36–45, Jan. 2013.[8] S. Narayanan and P. G. Georgiou, “Behavioral sig-nal processing: Deriving human behavioral infor-matics from speech and language,”
Proc. IEEE , vol.101, no. 5, pp. 1203–1233, Feb. 2013.[9] A. Vempaty, L. R. Varshney, G. J. Koop, A. H.Criss, and P. K. Varshney, “Decision fusion by peo-ple: Experiments, models, and sociotechnical sys-tem design,” in
Proc. 3rd IEEE Global Conf. SignalInf. Process. (GlobalSIP 2015) , Dec. 2015, pp. 1–5.[10] C. A. Sims, “Implications of rational inattention,”
J. Monet. Econ. , vol. 50, no. 3, pp. 665–690, Apr.2003.[11] C. N. White, R. Ratcliff, M. W. Vasey, and G. McK-oon, “Anxiety enhances threat processing withoutcompetition among multiple inputs: A diffusionmodel analysis,”
Emotion , vol. 10, no. 5, pp. 662–677, Oct. 2010.[12] R. Ratcliff and H. P. A. van Dongen, “Diffusionmodel for one-choice reaction-time tasks and thecognitive effects of sleep deprivation,”
Proc. Natl.Acad. Sci. U.S.A. , vol. 108, no. 27, pp. 11 285–11 290, Jul. 2011. [13] A. Vempaty, L. R. Varshney, and P. K. Varshney,“Reliable crowdsourcing for multi-class labelingusing coding theory,”
IEEE J. Sel. Topics SignalProcess. , vol. 8, no. 4, pp. 667–679, Aug. 2014.[14] A. Vempaty, Y. S. Han, and P. K. Varshney, “Targetlocalization in wireless sensor networks using errorcorrecting codes,”
IEEE Trans. Inf. Theory , vol. 60,no. 1, pp. 697–712, Jan. 2014.[15] B. Kailkhura, A. Vempaty, and P. K. Varshney,“Distributed inference in tree networks using cod-ing theory,”
IEEE Trans. Signal Process. , vol. 63,no. 14, pp. 3715–3726, Jul. 15 2015.[16] J. R. Anderson, C. F. Boyle, and B. J. Reiser, “In-telligent tutoring systems,”
Science , vol. 228, no.4698, pp. 456–462, Apr. 1985.[17] A. T. Corbett, K. R. Koedinger, and J. R. Ander-son, “Intelligent tutoring systems,” in
Handbook ofHuman-Computer Interaction , M. Helander, T. K.Landauer, and P. Prabhu, Eds. Elsevier Science B.V., 1997, pp. 849–874.[18] H. S. Nwana, “Intelligent tutoring systems: anoverview,”
Artif. Intell. Rev. , vol. 4, no. 4, pp. 251–277, 1990.[19] R. Nkambou, R. Mizoguchi, and J. Bourdeau,
Ad-vances in intelligent tutoring systems . Heidelberg:Springer Science & Business Media, 2010.[20] J. Levinson, J. Askeland, J. Becker, J. Dolson,D. Held, S. Kammel, J. Z. Kolter, D. Langer,O. Pink, V. Pratt, M. Sokolsky, G. Stanek,D. Stavens, A. Teichman, M. Werling, andS. Thrun, “Towards fully autonomous driving: Sys-tems and algorithms,” in
Proc. 2011 IEEE Intell.Veh. Symp. (IV) , Jun. 2011, pp. 163–168.[21] R. Vivacqua, R. Vassallo, and F. Martins, “A lowcost sensors approach for accurate vehicle localiza-tion and autonomous driving application,”
Sensors ,vol. 17, no. 10, pp. 1–33, Oct. 2017.[22] W. Raghupathi and V. Raghupathi, “Big data ana-lytics in healthcare: promise and potential,”
HealthInf. Science Syst. , vol. 2, no. 1, pp. 1–10, Feb. 2014.[23] Y. Gil, “Thoughtful artificial intelligence: Forging anew partnership for data science and scientific dis-covery,”