Summary: This post discusses one of the verification implications of the next 20 years of automation on verification: The problem of trying to predict how people will react to change.
A previous post (and another one) discussed the next 20 years of automation and their general implications. This follow-up post discusses one of the implications for verification / simulation / modeling. If you are not interested in verification (i.e. the process of finding bugs / misbehaviors before they annoy or kill somebody) now would be a perfect time to disembark.
Still here? OK.
I decided there are too many implications, so I’ll be spreading those over multiple posts. This is one. I already discussed some of the implications in these two AV-related posts: The need for a lot of virtual simulation, the push for transparent verification, and the need for big scenario catalogs.
So here is another important implication. Let’s first review this future: There will be (say) a 20-year process, bringing about: Lots of new automation; Huge turmoil in various industries; Lots of unemployment, with something like Universal Basic Income; A lot of occasionally-employed, highly-educated, Video-game-savvy and VR-savvy people with a thirst for meaning; A video-game industry integrated with movies, television, education and social networks.
This seems perfect for:
Better ways to predict future behavior
When you try to simulate/model future systems (e.g. to answer questions like “how much will a new train line help this city”), one of the big issues is that people are pretty bad at guessing how such external changes will really change their behavior. I have discussed here and here how hard it is to predict second-level implications of changes (the “bridge to Malmo” issue), and the problem of “stated preferences vs. observed preferences”.
However, in this future world, I think this will become significantly easier, since the following starts to make sense:
- Create an immersive, Virtual-Reality-based city-simulation. In fact, create two variants of that: One before and one after the train line got built.
- Get a bunch of these partially-employed people (of the appropriate demographics / culture etc.), Mechanical-Turk-style, to “live” first in the “before” variant and then in the “after” variant, and observe their behavior
- Repeat several times while varying various assumptions and conditions
- Teach artificial agents the behavior you learned so far, and drop them into the simulation instead of the humans.
- Do massive further simulations, varying even more parameters. Now you can really investigate the optimal parameters of your new train line.
This idea (which I initially got from my friends Nadav Levy and Itzhak Benenson) seems pretty reasonable to me. Much depends on the ability to create convincing simulated worlds fairly cheaply, which brings us to:
The influence of VR and video games
Note that VR equipment is an important part of the above story. VR lets you “really be there”, and thus makes the whole simulation much more realistic. This improves the chance that people will discover “what they’ll really do after the change”. Of course, you have to run this several times with several assumptions to get reasonable coverage.
VR is already starting to be used for checking “how some proposed setup will feel” (e.g. in the healthcare construction industry). One thing that is missing in current VR sets is the ability to fully track user movements.
BTW, my intuition is that VR will make a big change (mostly for the better) in our lives. See Why Should You Care About Virtual Reality? Because It’s A Source Of Hope and the (anecdotal) Virtual Reality Aimed At The Elderly Finds New Fans.
A related issue is that I expect the mighty video games industry (now bolstered by VR) to make huge gains in simulation and visualization technology. Video game technology will probably sort-of merge with movies, education and so on, and will become even more indispensable for people’s life (now that many less people will be working).
You may or may not like the fact that more and more people will spend much of their time in video games, but I think it will happen. And one benefit will be the creation of improved toolkits for defining simulated worlds, scenarios and so on. This will be a big enabler for verification: Not only will it help the “prediction” task discussed above, but also the task of verifying Intelligent Autonomous Systems in general.
For instance, AV manufacturers / regulators will have an easier time creating realistic simulations containing AVs, streets, people and so on. I described here what we (verification professionals) can already learn from advanced game design languages.
Finally, even with excellent VR and excellent video games infrastructure, this will be non-trivial.
Trying to assess how people will behave gets harder with longer time horizons: Using people-with-VR as drivers and pedestrians in an AV simulation (to extract typical / interesting behavior) seems not so hard.
But using people-with-VR in some life simulation to assess what-happens-if-we-construct-a-bridge is tougher: You need good, exploratory simulation (which is user-driven, like some video games). You also need users who are willing to play this for extended time periods (and get tired / hungry during the play). Assuming that the future will arrive at a faster and faster rate, such predictive simulation will become both indispensable and more challenging.
I’d like to thank Sandeep Desai, Amiram Yehudai and Ziv Binyamini for reading previous versions of this post.
2 thoughts on “Future of verification: Better ways to predict behavior”
Using VR to predict future behavior is going to have significant growing pains. The behavior of people online is notoriously different than their behavior offline. VR is no different and a lot of care will need to be taken to enforce “reality”. Some sort of risk/reward system will most likely be required to enforce normal behavior and even then, statistics will need to be used to eliminate outliers.
All of that said, as the cost of VR goes down, and VR engines become more robust (larger libraries of real world places, etc…), all of this is inevitable.
I agree. Continuing this thread, perhaps quite often it will not be the VR part, but rather the convincing-simulation part which will make a difference. Note that I assume that part will also improve a lot in the future.
Got a related comment from Nadav Levy, who said:
“In the above context VR is a big leap foreword indeed, but “Participatory Simulations” can be utilized and provide important input/validation without the need for immersive surrounding, specifically when human decision making are unrelated to the surrounding, i,e. stock exchange desertions, real estate development, banking etc’.
What i mean is that as we live in a world that more and more decision are made through a computer interface the immersive part is only needed for some (perhaps the more interesting or hard) questions, while others can be tested using simple multiplayer computer interface games that are much easier to build. :)”