AI, Dark Swans, Google Map Directions, and Moore's Law

Some further thoughts on the Google map episode I described yesterday.

It strikes me that AI in general will have problems with "dark swan" events.  That's true by definition--if AI is trained by using a big database of past information, it can't be trained to handle events which aren't reflected in the database.  In many cases, like autonomous cars, the algorithms can be set to do nothing.  It would be similar to the "dead man switch" found on locomotives--if there's no longer an intelligence at the controls which can react properly to events, you stop the engine.  But in other cases it may be harder to define such cases.

In cases like I encountered yesterday there may be salvation found in Moore's Law.  Presumably Google samples current traffic volumes using some priority rules--sample most often the most traveled routes.  But priorities are needed only when resources are scarce; as technology becomes cheaper it would be possible to sample everything all the time (which is almost what the human sensory system does).

Victims of Google Map Directions

Had an interesting experience in NY last Saturday, which I blame on Google Maps and the limitations of AI.

We drove from Kingston where our B&B is located to Rhinebeck, where the NY Sheep and Wool Festival was taking place on the Dutchess County fairgrounds.  That meant crossing the Hudson and taking 2-lane rural roads.  I've an off-and-on experience with Google Maps--this morning it was on.  So the friendly voice advised me to take a right at the first light after crossing the bridge.  While my usual route would have continued on to the second light, I decided to cede my judgment to the Google gods. 

All went well--we drove a few miles (maybe 8 or so) over a series of winding roads but with no traffic.  Came out on Rte 9 just south of the fairgrounds but was able with a break in traffic to hand a left and get to the fairground.

Evening came and we're ready to return to Kingston.  Each year they seem to organize the exit from the parking areas differently.  This year I ended up heading south, not north, on Rte 9.  Traffic was stop and go, mostly all stop very little go.  Somehow I had to head north to the bridge.  The Google voice advised me to take a right, retracing our route of the morning going the opposite direction.  I did.  Big mistake. 

I think what the Google algorithm must do is periodically sample the times on alternative routes, and recommend the fastest.  I suspect in areas such as we were in, they don't sample very often.  Consequently, maybe at 4 pm the alternative route was marginally better than the main Rte 9 north.  But the algorithm kept sending cars that way.  The problem was likely not only the winding roads, but the light where the route met the road to the bridge.  Since the big volume of traffic was on the main road (199 I think), the traffic light favored that, only permitting two or three cars at a time from the alternative to come onto the bridge road.  The end result: immobility.  At about the 1:30 mark I yielded to the advice of my better half, found a way to do a u-turn, and went back to Rte 9, which turned out at about 7 pm to be almost empty. 

The problem IMHO is Google couldn't keep track of how many cars it had directed the alternative way compared to the carrying capacity of the road.