FROM THE EDItOR «

Research Cycles

early all goes in cycles, This was followed by another AI these, it was only recently that I saw it and I have already experienced “chill” in the 1980s [4], but the field is codified in the form of a “hype cycle” N this many times in my career. For now undergoing a resurgence given [8]. The typical curve, shown in Figure 1, example, when I started on the fac- the excitement about 1) self-driving ve- plots expectations (sometimes labeled ulty at Stanford in 1994, it was at the hicles (SDVs), building upon the many as visibility) versus time and has the end of the period of intense interest years of research into robotics and the characteristic shape of a transient re- in controls-structures interaction [1], control of embedded systems, and sponse of a lightly damped system to which had lasted the entire time I was 2) deep learning, which builds on the a step input, but with the longer tim- a graduate student and postdoctoral earlier work on neural networks and escale response dominated by that of candidate (a period of approximately reinforcement learning (see [5]–[7] for a system with a slow (stable) real pole. eight years). However, by the time the technical details and a historical per- With illustrative names, such as peak Shuttle experiment we designed on the spective on that approach). of inflated expectations, trough of disil- use of active control for space struc- There are also many new exciting lusionment, and plateau of productivity, tures flew on STS-67 in March 1995 [2], areas in the control field, such as the the curve shows the tendency for early that interest had waned significantly control of cyberphysical systems using expectations to become “overblown,“ and the funding agencies had largely techniques including linear and signal which, once realized, leads to negative moved on to different topics. I have also temporal logic; the control of biological press and a large overcorrection dur- seen interest in many other research and medical systems using techniques ing which time many investors and re- threads come and go, such as research such as model predictive control; and searchers tend to switch their focus to in the 1980s–1990s into robust control the planning and control of networked other problems, leaving the remaining (H∞ and μ), neural networks, and fuzzy teams of unmanned systems (collec- few to resurrect what is left. Reflecting logic and research in the 1990s–2000s tively called UxVs) for commercial, on that cycle led to the recent comment into formation-flying spacecraft. farming, and environmental applica- about the AI field [4] “‘There’s definite- Of course, similar cycles occur in oth- tions using a variety of techniques, in- ly hype,’ adds Ng, ‘but I think there’s er fields, with perhaps the most famous cluding swarming. such a strong underlying driver of real being the artificial intelligence (AI) While I was aware of the typical cy- value that it won’t crash like it did in “winter” in the late 1970s following the cle of research for such as previous years.’” Lighthill report that, in a section titled “Past Disappointments” in [3], reported Most workers in AI research and in related fields confess to a pronounced feeling of disap- pointment in what has been achieved in the past twenty-five years. Workers entered the field around 1950, and even around 1960, with high hopes that are s

very far from having been realised Expectations on i

in 1972. In no part of the field have n

the discoveries made so far pro- er ectat gg p i x duced the major impact that was r TroughTrough ofof SSlopelope of PPlateaulateau ooff Trigger T Expectations E Peak of Inflated Peak of Inflated then promised. Innovatio DDisillusionmentisillusionment EEnlightenmentnlightenment PProductivityroductivity Time

Digital Object Identifier 10.1109/MCS.2017.2697199 Date of publication: 18 July 2017 FIGURE 1 An example of a typical hype cycle plot [8]–[10].

AUGUST 2017 « IEEE CONTROL SYSTEMS MAGAZINE 3 John Valasek (right) thanking Jonathan How after his seminar at Texas A&M University­ Jonathan How below his favorite X-plane during a recent visit to the National Museum of during the celebration of the 2017 Texas the U.S. Air Force in Dayton, Ohio. Systems Day.

The downturn after the peak in the the ones analyzed have made it past the As always, I look forward to your expectations can occur for many rea- trough of disillusionment (with virtual feedback on this topic. sons and could be driven by technical reality being the primary exception). (too hard), legal (not allowed), policy It is important to recognize that the References (might be allowed but has undesirable amount of hype about a can [1] J. R. Newsom, W. E. Layman, H. B. Waites, and R. J. side effects), and/or commercial (not have a significant impact on the type of Hayduk. (1990 Oct.). The NASA controls-structures­ interaction technology program, NASA-TM-102752. cost-effective) factors. For scientific research being done by researchers in [Online]. Available: https://ntrs.nasa.gov/search endeavors, we tend to focus on the that community because hype has a ten- .jsp?R=19910006744 technical issues, such as whether the dency to dictate what is “valued,” and [2] D. Miller, J. P. How, K. Liu, M. Campbell, R. Glaese, S. Grocott, and T. Tuttle, “Flight results from algorithm is computationally tractable not necessarily in a good way. Further- the middeck active control experiment (MACE),” for a realistically sized problem, are more, while increased hype can sim- AIAA J., vol. 36, no. 3, pp. 432–440, Mar. 1998. the conditions on the stability theorem plify the process of obtaining funding, [3] J. Lighthill. (1973). Artificial intelligence: A paper symposium. London: Science Research too tight to be practically useful, or do those funds will also attract many other Council. [Online]. Available: http://www.chilton- the performance improvements of the researchers, thereby typically making it computing.org.uk/inf/literature/reports/lighthill_ proposed approach meet prior expec- more difficult to differentiate your work report/p001.htm [4] W. Knight. (2016, Dec. 7). Intelligent machines: tations. However, the policy and legal from others and/or make a unique con- AI Winter isn’t coming. MIT Tech. Rev. [Online]. factors often strongly depend on the tribution to the field. Available: https://www.technologyreview.com/s/ technical issues, as is the case for the While it is sometimes necessary to 603062/ai-winter-isnt-coming/ [5] J. Schmidhuber, “Deep learning in neural net- safety analysis of SDVs and UxVs, so it “enhance expectations” to stoke interest works: An overview,” Neural Network, vol. 61, pp. is important to be aware of, and help in the ideas and work, it is also important 85–117, 2015. doi: 10.1016/j.neunet.2014.09.003. address, those issues as well. to perform the research that addresses the [6] L. Deng. (2014). A tutorial survey of architec- tures, algorithms, and applications for deep learn- Gartner publishes a yearly sum- issues that might lead to a down turn. If ing, APSIPA Trans. Signal Inform. Process., 3. [On- mary of its Hype Cycle [9], which is de- successful, work in that direction would line]. Available: http://journals.cambridge.org/ signed to provide a broad perspective be ahead of the pack, and even if not, then abstract_S2048770313000097 [7] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learn- on technologies and trends that have the results might provide the necessary ing,” Nature, vol. 521, no. 7553, pp. 436–444, 2015. high potential impact. Perhaps not too damping on the hype peak overshoot. [8] Hype cycle on Wikipedia. (2017). [Online]. Avail- surprisingly, items such as SDVs, the In- While there are some criticisms about able: https://en.wikipedia.org/wikiHype_cycle [9] Gartner’s 2016 hype cycle for emerging tech- ternet of Things, unmanned aerial vehi- the hype cycle, including that there may nologies identifies three key trends that organiza- cles, and reinforcement/deep learning be too much hype about it, I think this tions must track to gain competitive advantage. appear prominently near the peak of is a useful visualization of the typical [Online]. Available: http://www.gartner.com/ newsroom/id/3412017 the 2016 curve. An analysis of the recent ebb and flow of interest and research [10] H. Katz. (2016. Aug 31). Gartner’s emerging (2013–2016) of these hype cycles funding. As such, there is a fundamental technology hype cycle—Updated for 2016! [On- [10] is insightful primarily in that it question that I think should be carefully line]. Available: https://whatitallboilsdownto .wordpress.com/tag/gartner/ makes it clear that 1) many technologies considered before investing time/effort relevant to the IEEE Control Systems So- into a new area (such as when making Jonathan P. How ciety community have hovered near the career decisions)—where on the hype peak for some time and 2) very few of curve is this field of interest?

4 IEEE CONTROL SYSTEMS MAGAZINE » AUGUST 2017