-   EcoModding Central (
-   -   IBM's Watson applied to the Ecomodder (

SoobieOut 02-22-2011 04:50 PM

IBM's Watson applied to the Ecomodder
After seeing last weeks crushing Jeopardy defeat of the best Human player at the byte of Watson, IBM's latest supercomputer, I wondered if the same Machine learning concept could be applied to saving fuel.

There is an excellent NOVA show on how Watson was developed, seems it lost all the time until they applied machine learning from the past 100 episodes of Jeopardy.

With this in mind, could it be possible to have a computer learn from a professional ecomodder in the realm of Engine-off coasting (EOC). Then use that learning in future chip programs?

I am not talking about just PID loop tuning, more in the thinking of pattern recognition adaptive tuning.

Has this been tried before?

ChrstphrR 02-22-2011 09:34 PM

Haven't tried it yet, though, the concept is very interesting.

I think having to power a rack of servers toted in a trailer would negate all the efficiency gains. :)

Seriously now, I think the first stages of this would have to be done first: hordes of data collection, so that there's something to find patterns within. Off the top of my head, I wouldn't know what things you'd want to track and log first.

NHRABill 02-22-2011 09:54 PM

I like your idea but would like there to be more research in designing a system that would detect the road pattern in front of your car. i believe the Watson program might be not a good fit but maybe some of the research that they are doing on driverless cars can be adapted to a driver assist program that worked like a "Intelligent cruise control".

The system would auto hypermill for the driver according to GPS and in car sensors such as lateral movement and weight being placed on each wheel. all of the data is currently being collected someone needs to organize it and put it together in a usable format...

Times like this I wish I was a programmer.

Frank Lee 02-22-2011 10:07 PM

I don't see this as having any chance of working unless it is not an onboard system controlling individual cars, but a system that coordinates between groups of vehicles.

RobertSmalls 02-23-2011 08:20 AM

In theory, a computer would know its road load and BSFC at all times, and know exactly where the "sweet spot" is located, and time lights perfectly 100% of the time, etc. In theory, a computer would run circles around the best human hypermilers. Instead of emulating us, it would run a huge optimization calculation seeking out the lowest fuel consumption solution that meets the driver's requirements (getting to work in x minutes, e.g.).

In practice, an AI can't get a driver's license yet, so it's moot.

vacationtime247 02-23-2011 09:44 AM

Like the idea. The use of a 'smart cruise' could be feasible. Say you traveled the same route to work each day. After driving back and forth a few times the computer would learn the route and road. Really like the idea of this being used on an electric car. Could extend the range further with P & G and EOC at pre-determined locations en route.

jamesqf 02-23-2011 12:32 PM

In fact, IBM did some work on something similar a few years ago, for hybrids. The processor would look at the route & conditions, and figure out the best mix of electric & gas for economy.

Arragonis 02-23-2011 05:39 PM

Computers are stupid, I know I program them :D

Seriously though think about your last journey hypermiling. You spotted a hill ahead, maybe up or down. How did you know ? How could you describe it so that something that has no idea what a hill is would know ? And then how could you tell this thing what a hill looks like ?

Colour ? Naw some have snow, some are concrete, some are grass, some are desert...

Is it above your horizon or below ? Naw - you have smaller hills in front of you that you probably ignore.

Are we tipping up or down ? Well a computer can do that but then when does the hill end as that affects how long we cruise or how long we 'pulse'.

How do we ignore the cars in front so we just look at the hill.

Actually should we look at the cars because there is a hazard ahead - maybe a junction, maybe a busy spot or maybe just a numpty on a mobile.

We (humans) would spot all of this and react. Maybe we would react differently but we would react in the way that our (and others we know) experiences have taught us.

However at this point the computer is still working out what a hill is.

Frank Lee 02-23-2011 05:42 PM

I said it won't go because a lot of our driving adjustments are related to other traffic, perhaps even moreso than to terrain. Having a Watson doesn't give our car a bird's eye view of the traffic.

RobertSmalls 02-23-2011 07:12 PM

Nah, the computer knows all about hills. It has GPS, and Google Maps. It knows the exact geometry of every hill on every possible route from here to your destination. It also knows whether there's a stop sign at the bottom of the next hill, and exactly how fast to crest the hill in order to regen to a stop (accounting for crosswinds, even).

It can also see other cars and how fast they're going. Google has been running autonomous cars on public roads for about a year now, using cameras to keep an eye on traffic, signals, etc.

What a computer can't (yet) do is stereotype. Folks leaving a fast food restaurant are likely to be driving while distracted. A beige Buick is likely to go below the posted speed limit. National Fuel trucks like to drive PSL+15, and get aggressive if you don't. And so on.

You can write down your ScanGauge trip averages every day on your way to work, and try to get an idea of what routes and techniques pay better than others. But a computer can do so from red light to red light, with enormously impressive precision and memory. Humans have the edge for now in pattern recognition, but we can't hold a candle to an expert computer system in terms of memory, processing power, and skill with optimization problems like fuel economy.

Without a doubt, the best fuel saving device that will be developed this decade is the AI autopilot.

All times are GMT -4. The time now is 11:15 AM.

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2022, vBulletin Solutions Inc.
Content Relevant URLs by vBSEO 3.5.2
All content copyright