Neural Networks In The Home?

Hougaard asks: "I'm investigating the use of neural networks in homes for a architectual project. Do you have a neural network that controls the light or the temperature in your home? Do you want one? What other features would you like to have in a real intelligent home? How should you learn and train the network? A remote with a 'Good' and 'Bad' buttons or something completly different? The only real life example I have found is the use of neural networks in elevators to predict elevator traffic. Yes I know that Cisco and others have the 'Internet House' - but that is just a home with a network - not really a revolution - and not very intelligent."

Comments

-57 Celsius!

[Anonymous Coward]

New record low temperatures in Siberia!

-57 Celsius (-70.6 Fahrenheit)

Wrong Tool

[Masem]

I wouldn't suggest neural networks for controlling house lights, save for the possibility of heating/cooling systems.

Neural networks are meant to try to 'black box' complex behavior into a nice package, with the ability to learn and adapt from past and present behavior. While the running of a household is certainly a complex pattern, it's TOO complex with too many variables to be effective. NNs can also over or under predict, and I'm sure there are things that you do not want to go on or off without your input.

A NN *could* help with efficient heating and cooling of your house, of course; you'd need to have a large data set of times, indoor and outdoor temps along with windspeeds (or wind chill factors) for about a year to get an effective yearly cycle in place, then with the time and conditions, it should be easy to have the network control the temperature and minimize energy use while maintaining a comfortable room. This is because all those factors can be realized as a complex yet predictable pattern (as there is no human element involved), and thus a trained NN should be able to do the trick.

What you'd probably want for the rest of the house is the use of a rules-based logic or fuzzy logic. This is more adaptable to individual users (specifying their own rules), and it is more than adequet for house control.

Utility of Neural Nets

[MenTaLguY]

The basic utility of neural networks is basically the same as "genetic programming" .. they avoid the necessity of having to describe a solution to a complex (or merely subtle) problem.

The trained network or the evolved program solves your problem, but it's basically "black box magic" -- you never really will understand completely how it works or how it was arrived at. Both can also exhibit unexpected properties.

They're overkill for most problems where you CAN describe an adequate "deterministic" solution. Neural nets also require a fair amount of overhead, too.

Re:Utility of Neural Nets

[Flavio]

The basic utility of neural networks is basically the same as "genetic programming" .. they avoid the necessity of having to describe a solution to a complex (or merely subtle) problem.


I know, but I believe the hassles one goes through in the learning process aren't worth the final result, specially in a very deterministic environment like a house. As another reader pointed out, you wouldn't like the net to deal with events like waking up at 4 am for some random reason, a fact that isn't likely to repeat itself.

Flavio

Re:Two issues

[maggard]

Generally most folks don't consider X-10's to be "Smart Houses". An argument could well be made for them, particularly the more elaborate setups but in most cases what's referred to as a "Smart House" is more then the extended lightswitch-type devices X-10 modules usually are.

Sensing, monitoring, complex-responses, etc. would be more in line with what is often meant. Your definition may of course vary widely.

That said are you aware of anyone doing any adaptive/learning/non-declarative stuff with these rudimentary devices?

It's been done

[gary.flake]

Michael Mozer of U. Colorado CS department wired his entire house with sensors and controls connected to neural networks and other machine learning systems. He did this at least five years ago, so the idea is hardly new. The house has a web page with an overview and another link that shows the status of the house.

One great story about the house has it that Mozer's students would call his house whenever the toilet sensor showed that he was sitting on the can for more than a minute.

I believe that particular sensor was later disabled.

--GWF

Re:MIT MediaLab

[xyzzy]

This is the right idea, but it doesn't require a neural network to do it for you. It just requires a computer that can tell when you are awake or asleep!

The thing about neural networks is that they are good for making fuzzy judgements about patterns based on trained input. But I KNOW the way I like my lights to be, what temperature I want and when, and there aren't too many possibilities! Once I program everything into the thermostat or X10, I'm likely to be fine -- I'm not sure what a neural network would DO in this case.

For the media lab idea, you COULD use an NN whose inputs are things like the room you are in and some notion of the amount of movement you have made in the past few minutes, but you need to be careful here -- you don't want it turning the lights off when you are sitting on the couch quietly reading.

A REMOTE??

[meridian]

Dont you think a realy intelligent house would be completely voice activated for those who could speak anyways. Id prefer to scream at my walls that they are stupid idiots for turning on the dishwasher in tthe middle of the night instead of having to fumble around for a remote control and figure if im telling it its good or bad. I guess having a battery powered system would be an important feature, put good use to an old 486 laptop that has no other use than to also control the houses security system. And Solar Power would be a good feature to build in as well.

Re:A REMOTE??

[roman_mir]

A house of this cost would have much wider water pipes than the standard american pipes and the water pressure would not be a problem.

Slight aside.. but AI related.

[GC]

I always thought that once computers have learnt to recognise when we appriciate their actions they will become hell of a lot more helpful to us.

Fundamentally Doomed to Failure

[Somnus]

There is a serious "engineering tradeoff" with neural networks: You are exchanging reliability and precision for creativity and flexibility. How much unreliability and imprecision are you willing to stand in your home automation?

Imagine if your goldfish was running the lights in your family room ...


*** Proven iconoclast, aspiring epicurean ***

House Training

[Raylen]

The way the house knows it did something "Bad" could be when the user intervenes. For example, someone turns a light on at 5:30-ish pm for a few days it figures out "This person wants a light on this time of day."

The only real problem with this is how do you collect information to base these conditionals on. The example above used only time. Time, time of day, seasons, etc are pretty easy but do you want cameras following you around the house or even pressure sensors? When someone walks into the room, light up except between the hours of 8am to 6pm (varying on time of year) when the giant window is collecting lots of light.

Anyway, my point that got lost somewhere up there is the more information you have, the better (hopefully) the system can predict wants and needs. However how much information do we want these systems knowing?

Re:i hate to say this but

[lizrd]

The problem with carrying around a badge around at home is that I wouldn't do it. My patterns at home are vastly different than when I'm at work or school. When I'm at work it can be assumed that I'll be doing various things that make the carrying of tracking devices fairly unobtrusive, however most of the ways of tracking this assume that I am wearing pants. This is a very reasonable assumption at work when I am at work I do wear pants and the accompanying wallet, key ring, cell phone and PDA. However, this all changes when I get home. The wallet, key ring, cell phone and PDA get deposited on the coffee table and the pants with pockets get dropped in the hamper then I put on cutoff sweat pants (no pockets and eat dinner, watch TV, surf the net, talk on the phone or whatever I'm planning on doing that evening. How is the system going to effectively track what room I'm in without the items that I customarily carry? Is it going to be confused that my shoes (yeah, just like in Enemy of the State) are on the mat by the door, the wallet/cellphone/PDA pile is one the coffee table, the pants/belt is in the hamper, the watch is behind the kitchen sink (took it off to wash dishes)? Rest assured that I am not standing by the front door for hours, nor am I standing on the coffee table or hiding in the hamper or swimming in the kitchen sink. Well, come to think of it, I might do those things but it's not the normal case.

I don't object to carrying identifying items with me when I go out, but I just don't do it at home.
_____________

Possible house uses

[mOdQuArK!]

People have already mentioned some of these, but I'll reiterate anyway :)

1. Trainable room temperature, based on temperature outside & current temperature inside, time of day/year, and whether anybody is in the house or not (and possibly WHO is in the house), vents to individual rooms, cost. Simple training using remote to indicate currently where too hot or too cold & measurement of fuel/electricity use, with some factors to make the neural net understand that temperature change is not instant.

2. Trainable lighting - based on light coming in from outside and/or time of day/year, and whether somebody is in the house/room (and possibly who is in the house/room) and whether or not there are other independently-controlled sources of light, & electricity cost. Possibly hook up with automated window blinds, to control light from outside (perhaps also a privacy factor). Similar training mechanism as for heating/cooling (too bright/too dark/electricity cost).

3. Trainable water temperature & amount (depends on which faucet (kitchen/bathroom/shower/tub/etc), who is asking, time of day/year, competition for water sources, how much water is left in the hot water heater, cost. Train by hot/cold/too much/too little/cost.

4. Home entertainment (at least volume for stereo, based on where listener is located & time of day). Could also be choice of station, both for radio & TV, although I'm sure the neural net would end up getting "punished" regularly :)

5. Security - choose reaction (turn on more sensors, turn on fake dog barking, turn on surveillance, turn on siren, alert police, alert owner, etc) based on which combinations of sensors are triggered. Send the robot with the gun? Training this net could be fun :)

I'm sure people will be able to think of plenty more - question is, will they be simple enough to implement?

Re:i hate to say this but

[}{avoc]

Bill Gates' book, I belive it was called "The Road Ahead", which was published quite a few years back (around the Win 3.11 days) contained a CD with a "virtual tour" of the future Gate's mansion which outlined the various innovations the home would have, which I must say, were rather impressive. The things I can remember right off the top of my head were a tracking badge you would clip on when you enter the house, which would allow music to follow you around, art (displayed on wall-mounted flatscreen monitors) would change to match things you like, and phone calls would be routed so that only the phone nearest to you would ring. I'm not sure, but I think I recall it mentioning temps and lights adjusting to fit your prefrences, though I am not sure... If anybody wants, I'll dig up the CD and see if I can get more specific information.
-Dan

Re:Neural Networks can't learn.

[Kepps98]

Back propogation isn't the only way to teach a neural net. There's plenty of ways you can set up the system to teach itself a function. Temporal difference learning is one such way that uses back propogation to manipulate the network weights but generates its own internal error signal without any external teacher. It takes a lot longer to learn the function but it can be done.

Re:Neural Networks can't learn.

[Maurice]

They don't learn like humans do, but a (three layer) neural net can express *any* function. That is, if you have a finite number of combinations of inputs and outputs and you don't have the answers for each one, you can enter the ones you know and extrapolate from there. In this case there are not many possible combinations for the lights, etc. so the results should be reasonable. Neural nets have been successfully used for face and handwriitng recognition, which are hardly trivial. Neural nets are useful when you need results more or less immediately but the system can improve as you go along and get more examples.

Work at PARC

[Animats]

The Xerox PARC crowd fooled around with something like this a decade ago, but it wasn't very useful in office environments.

Adaptive room control would probably be most useful for conference rooms and lecture halls. Lighting, HVAC, blinds, sound systems, and microphones all need to be coordinated, and it's usually done badly unless there's an operator for all the stuff. There, you might have a system using motion detectors and maybe a TV camera to distinguish between "room is empty", "a few people are in the room", "room is full", and "room is overcrowded", along with "writing on board", "using projector", "sun streaming through windows onto screen", etc.

Most of the smarts is needed to figure out what the room situation is from the sensors. What to do is mostly a mode thing.

Neural Networks CAN learn!

[DaBs]

First, it is a misconception to think that the back propagation-model is the only type of neural network, there are many more! There are also several types of models that do not need any 'external program' to modify there behaviour (for example Interactive activation).
Second, the fact that usually an algorithm is being used in backprop is not so much because the model can't do without, but because this is the easiest way (for now) to implement this subtask. The idea op backpropagation IS part of the model and may or may not be part of some implementation..

cheers,
DaBs

hmmm

[nomadic]

Did this make anyone else think of the Ray Bradbury story, "There Shall Come Soft Rains", about an automated house that outlives its owners, and keeps operating as if they were there?
--

Don't instruct, let it observe

[www.sorehands.com]

Use the same controls that people are familiar with and allow the house to learn from corrections that are made.

The house would have heat control that considers the usage patterns and the outside weather. The house would detect motion to control which rooms are heated, lit, etc.

There's no reason not to

[drinkypoo]

Though I don't think a Neural net is necessary, or necessarily useful in this case.

There are lots of interim steps that you can take to make your house seem smarter. For example, at one well-publicized geekhouse in Santa Cruz, Darkwater (webpage seems to have gone missing), my friend Charlie (Creator of OmniRemote for Palm Devices (shameless plug) rigged up a system to make heating work for individual rooms. To wit, each room was fitted with its own digital thermostat, and a flap in the vent which was opened and closed via some sort of servo system. If any thermostat in the house wanted heat, then the heater kicked on (think of this as a big set of OR gates) and any room which wanted heat would open its vent.

That's just one example; There are probably lots of things which could benefit from similar improvements, or at least similar applications of geek brainyness.

If you DO decide to go the neural networks route (which is worth doing just for the geek/hack factor, IMO) then you're going to have to decide what criteria you want it to base its decisions on, how you're going to track all of that criteria, and what kind of sensors you're going to use for anything that is sense-based and not knowledge-based.

For example, one poster has already suggested that the system learn from human intervention. They use turning on a light at 5:30 as an example. So you go and replace all your light switches with X10 devices, and you have some device which logs your X10 codes and feeds them into the neural net. But now you have to decide which pieces of information to add; As you've probably already figured out, this is the key to having this work well. You should track the day of week, though also tracking weekday/non-weekday has a great potential to help this feature learn to be useful more quickly.

But that's not nearly enough data; There's also other useful things like "Am I present in the house", which might mean you want other devices turned on; You're going to have to have some default behaviors there. Or perhaps "is it Christmas morning", in which case you might want the lights already blazing before you get downstairs so that you don't go blind trying to adjust to the light levels before your morning coffee or penguin mints or whatever geeks consume for pep these days.

The other things you should make sure of are that you can still live your life without the system, and that it doesn't get too uppity. If you make a change in state, you don't want it changing it back before you're done with that state change.

And do yourself a favor, implement voice recognition, and make it make plain sense. Use an activation word or phrase; I recall some system (in a Niven book, maybe?) using "Prikazvyat" (sp?) which IIRC is Russian for "Command" or something. It's been a while. Giving the system a distinctive name is another fine idea; "Computer" is probably a poor choice. Don't worry about being able to issue commands in English, either. Having the computer recognize what you're saying 99% of the time is more important than flow of speech. Saying "Halloran, lights this room up full" is an acceptable compromise, and should be reasonably easy to recognize.

Yes, I know what Neural networks might do.

[WolfWithoutAClause]

Suppose you have a dimmer switch, temperature gauge for your house etc.

Wouldn't it be great if the neural network set the system to the right value BEFORE you asked it to?

So the way it works is that everytime you change a setting, it assumes that you've changed it for a reason, so it tries to correlate its sensory inputs (time, outside temperature, inside temperature, humidity, infrared sensing of people, time of year, how dark it is outside etc. etc.) with what changes you've made.

Then next time when its sensory state changes in that way it sets the outputs that way before you ask it to.

Anyway that's the theory. And its a way cool theory in my opinion.

There are lots of problems. Still, not having any hot water for a cool high tech reason is better than not having any hot water because you forgot to turn on the water heater ;-)

Neural Networks vs. Stupid Networks

[Ace905]

The question posed in this article really doesn't make much sense. Why do we want to see neural networks in our homes if we don't have an end-goal?

I think most people live in such a way that intelligent-guessing of there habits is going to be more of a problem than simply giving them control of there environment through a 'stupid' network.

For example, as many posters have mentioned here; the lights in your home. You can't *guess* when a person wants a light in a room on based on any number of variables, just the same as you can't go in to someone else's house and control there lighting for them without getting feedback from them; ie: "I want lights to turn on in all rooms I *may* enter, a light on in the room I am in (obviously) and lights to turn off behind me, in rooms I can not enter directly from the spot I am in". Well this type of processing requires only sensors; deciding on whether the person wants bright, medium or dim lights is based completely on how they feel at the moment and no neural network has the ability to predict that with even effective accuracy.

Other examples of house-hold tasks that do not require a neural network: Toilet Flushing Heat (stupid networks control heat better) Water from taps Bathtubs filling at specific times Windows automatically tinting for heat-retention Dishwashing Laundry

etc. A neural network is meant to bridge a gap between people's simple requirements, and complicated digital processing. There is no requirement for digital decisions that are literally, "on" or "off".

intelligent houses

[Alien54]

I can see a house that is intelligent. Star Trek, and all that. That could be useful, fun, etc.

But what I do not want is the intermediate levels of intelligence.

For example, would I want a house as intelligent as a puppy or a parrot?

This gets into matters of personality, etc. Most usable modules would likely not have any viable personality, because otherwise we would have to much independance and possible arguements, and other things that are quite maddening.

I do not want to live in the 21st century equivalent of a Monty Python skit.

All this aside, I would see that a nueral net would be useful for learning to automatically adjust things like heat and air conditioning. Security may be okay, but I've seen too many movies to trust them with things like weapons and other defensive measures.

;-)

'Smart Homes' bug me

[DeadVulcan]

They bug me because it seems like a gratuitous use of technology. Believe me, I'm all for technology, but I think we should use it where it fits, not go nuts-blind with it. I feel (just a feeling) that trying to make our homes too smart is rather wrong-headed.

For example:

Heating control. Personally, I don't want to walk into a room and be surprised that it's ten degrees colder (or ten degrees hotter) because the A.I. failed to predict that I'll go there. I would prefer to tell the computer not to condition the air in certain rooms, possibly at certain times of the day or days of the week. Good old fashioned programming (with possibly something simple like fuzzy logic) will suffice. The real problem is making an intuitive user interface that a layman can use to program it.

Lighting control. Again, I don't think this makes sense, because lighting is far more dynamic. I want the lights on where I am, not where the computer thinks I want to go. It's a control problem, pure and simple. Just install sensors that read where I am, and control the lights directly. (Of course, sometimes I might want to override it.)

Central music control. This one is the most dynamic of all. Not only does it depend on where I am, but how I feel. There are neural networks that can verge on reading facial expressions these days, but I would find it rather creepy and annoying if the house started piping in music to try and cheer me up, or try and fit my mood. My feeling is that music selection is a database access problem, not an A.I. problem. You want to be able to choose what you want quickly; you don't need a computer to try and predict or intuit.

At the heart of it, I think my problem with the whole idea is this: I want control over my house. The A.I. of today is not intelligent enough to be useful. It would need to have almost human intelligence to be useful, and I believe that by the time computers are that smart, we'll be able to plunk them into robotic bodies, bypassing this whole question.

Places where neural networks might be useful are in low-level control systems for complex machinery, like the furnace, the water heater, or the plumbing. These would not constitute neural network control of a house, but rather the underlying subsystem.

--

Re:Not a problem

[Luminous]

But what if I don't want to constantly monitor what my house is doing to 'teach it'. Sure if the TV turned on and I didn't want it on I could turn it off, but isn't that like having a three-year old running around that you have to constantly monitor. I can barely take care of my plants.

Re:Well..

[Luminous]

How would the owner know if the house has adapted?

Some Problems

[Luminous]

1. The initial learning curve for the house would be months. I prefer to spend my time doing other things than teaching my house what temperature I like my baths, what level of brightness I like the lights in my living room at 5:00pm compared to 10:00pm, etc. 2. Exceptions are even harder to get programmed in in a short period of time. I'd hate to plan a romantic Valentine's Day with a candlelit dinner to suddenly have the TV kick on to Babylon 5 and all the lights switch to blue.

No remote.

[scott1853]

If it's to be revolutionary, it needs heat sensors to detect your body temperature and adjust the A/C or heater according to the preset temperature you have specified. Of course that idea is well suited for an individual. Multiple family members may be a little complex, especially if they have different comfort zones. Maybe an ID tag of some sort, but it would be nice to track individuals in other ways.

For lighting, it would be nice to tell if you are coming from a room that has had the lights off, such as waking in the middle of the night. In that scenario, it would be nice to have lights come on automatically but be dim. Maybe if a person needs to be waking during the night hours for an early job, allowing the programming to gradually increase the brightness as the person wakes up would be a nice feature.

The real trick would be a simple way to do train the neural network. I'm not sure what would be best, but a simple method would definitely be a good idea, even if the individual only has to go through the process once, they won't use it if it's too complex to program. I know several people that have security systems but don't use them because they don't feel like learning the interface.

my take

[fleck_99_99]

I'd recommend using some sort of hybrid system of neural networks techniques and fuzzy logic. In fact, there's no reason not to regard a fuzzy logic system as a neural network...

And even those may be too exotic for what you're trying to accomplish. How many sensors and the like are you thinking of working with? Unless you plan to monitor enormous amounts of environmental data (temperature, humidity, light, and motion in every room at a bare minimum), your system is doomed to being relatively "stupid" anyway. But back to the point...

A fuzzy expert system would be an easy way to build the base of an "intelligent" house controller, by establishing simple variables and rules (a la "if KITCHEN-LIGHT-LEVEL is LOW then KITCHEN-LIGHT is ON-FULL" etc). This is easy to establish initially, and gives you an "out of box" working system. Some sort of feedback technique (a la back-propagation, but modified to fit what you're trying to accomplish) can be used then with the "good/bad" remote mentioned above. Basically, you have two things you can tweak iteratively when "training" the system in this kind of a case. You could lower the weight that certain rules have, or you can modify fuzzy variable definitions (what exactly is a "LOW" LIGHT-LEVEL? maybe it's even specifically, what is a LOW KITCHEN-LIGHT-LEVEL?).

I think the idea is very workable, although I'm not sure how terrific it'd work IRL. Make sure to include an automatic override! (No, dammit, I know I'm not moving. That's because I'm READING...)

"I'm sorry, I can't do that..."

That's what Rosie the Robot is for

[eclectro]

Forget neural networks, what people want is a Rosie the Robot for their homes.

A wonderful problem, a disappointing problem.

[mosch]

There are two problems with this discussion, one is wonderful. I'm making an attempt to identify them in the hopes that one or two more people might post intelligently and inspire some good ideas and responses.

The first problem is wonderful. The question that's been posed to us is vague, perhaps purposefully vague. It talks of "neural networks" without a stated problem. Is Hougaard looking for a way to run his heat and air conditioning? Does he want lights to automatically do what he'd like them to do? Does he want his house to predict what kind of music he'd likely enjoy at a certain point, and play it? This lack of definition seems to have stumped most respondants.

The second problem is a lack of informed, creative response. This is slashdot, a place where people like to think they have a better grasp on technology and it's implications than the rest of the world. This discussion is proving that it isn't so. Where's the creativity? Why are there posts which don't seem to indicate any hint of knowledge about how neural networks work?

I think this is a wonderful, amazing idea. Imagine if you will, starting small. Take one thing, for the first item, I'll pick air conditioning and heat. The system is set up such that in each room, there's a method of indicating if it is too hot, or too cold. The system could contain a clock, a basic weather station with access to humidity, temperature, wind speed and direction.

After a little bit, one could have a system which knows that if the wind has kicked up from the south, then the master bedroom starts to feel a little bit cold, but the rest of the house will be the same as usual. It will know that you tend to watch TV in the family room, which isn't over a basement, and thus gets a cold floor when the temperature has been below 50 for more than a few days. These are the kind of optimizations that would take forever to program into standard logic.

As a more interesting, harder to get right example, music. Let's say you have a centralized database of music, with automated access to your CD players, your mp3 database, and Music Choice on your satellite TV. That part's easy, I've got that. I also categorized all my music, no matter what the format is, and I'm sure I'm not the only one. After all, it's a hell of a lot easier to buy a stack of 200 CD changers, and write some software to control them, than it is to remember where it was you decided to file that Negativland cd, or that orchestral recording with Brahms and Rachmaninov on the same CD.

Now all we do is modify that software, to help the neural network understand what it is we're likely to want to hear at any given time. If I've been sitting at my unix box working (detectable by the state of xscreensaver), then I'm probably coding. Because I've manually picked the same genre of music every other time I've been activating these neurons, it puts on the new Dark Soho album, and I find myself listening to some slammin' trance. If it guessed wrong, well I just go to the standard music interface, tell it 'no, i wanted to listen to that midfield general ep i finally picked up' and eventually it figures out 'hey, this guy overrides his standard preferences, for stuff that's new to my database.' It might not get it right every time, but it'll likely be a hell of a lot better than what I'd get if I just put the player on random.

Surely somebody in this crowd of self-proclaimed geeks and cutting-edge thought, can realize the potential of this, and open your mind to the possibilities of a system like this. Possibilities like incredible profit. Ludicrously incredible profit.

Stop whining about what isn't possible, and think about what is! It's the 21st century, and if I can't have a rocket pack and a flying car, I want a house that plays dope tracks, automatically.

--
"Don't trolls get tired?"

Two issues

[maggard]

I think you're confusing two issues here:

1. "Smart Homes" that have intelligent appliances or other otherwise control their internal systems in a way more high-tech then the traditional discrete switches on the wall & the occasional independent light/water/etc-sensor & timers.
2. Neural Nets as a programming tool (method? algorithm? model?) vs. other more traditional explicit systems like cause/effect rules & preset series of grouped functions ("Party Mode", "Bedtime Mode", etc.)

Combining them is interesting but first you've got to find someone who has a "Smart House" then discover if they use "Neural Nets" in any way. Even on /. the number of folks who live in full-blown (or even partial) smart homes is incredibly small and with that small sample the odds of finding someone using neural nets is even more remote.

Consider posting on home.automation newsgroups, dedicated websites, and mailing-lists. There at least you'll have reached the first criteria of your study and can begin looking for the second.

I also wonder why neural nets would make any difference? Are you interested in systems that can 'learn' from an occupant? Are you looking to compare 'nets to other more traditional systems using sensors & statistics? Comparing different types of 'nets against each-other? Identifying learning-curves, ability to respond to differing situations, appropriateness of responses? How could one even evaluate "success" or "accuracy" against other automation systems? (Yes I know there are methods even for very fuzzy stuff like this but I can't imagine a sample-set out there being large enough to be meaningful.)

Not to be disparaging but unless your posting has been edited-to-idiocy it appears overly broad & extremely vague. Indeed what applicability any of this has to architecture escapes me. Systems Engineering, Electro-Mechanical Engineering, Artificial Intelligence, Human Interfaces: Yes but all of these are the provenance of specialists, not Architects (at least amongst the architectural curriculum I'm aware of.)

Or is this simply combining two hot buzzwords in a way to create a research project out of thin air?

I've got it

[nomadic]

This would make it great for pranks; sneak into someone's house and then train the network to play the theme from Psycho whenever someone takes a shower...
--

I don't think it would be a good idea

[Flavio]

Neural nets usually work well in areas where conventional binary logic fails. Text, image and speech recognition are examples.

A neural net in your home would, in my opinion, be too complex to bother with. Neural nets are self programmed through experience to perform choices that aren't easy to define in computer code.

For example, some time ago there was this slasdot post about a program which supposedly recognizes "innapropriate" images (i.e., porn). It can be used in web proxies, for example. The program didn't perform very well (in my opinion, it failed miserably, as the task is extremely difficult).

A very complex neural net could theoretically learn from experience to recognize the difference between a naked baby and two people having sex. You can't program such a net by hand because there are so many neurons involved and influencing other neurons in ways so complex you can't even imagine.

Now why would you want to use that to switch/dim lights in a house? The net could learn your behavior through experience at first, for example. It could have this "learn mode", much like smart network switches do. The net itself would be useless at this point, and it would record what you do at particular times of the day and perhaps make notes on your moods and act accordingly. After this learning period it would do things on itself and be corrected if the choice was incorrect. But hey, I really don't think you need the black magic that neurons give you to notice that.

The bottom line is that people are usually very predictable. Neural nets are great when nuances from an objective POV are fundamental (like position/density of beige colors in an image as an indicator of porn) and this doesn't seem to be the case.

Flavio

neural-network wants

[buss_error]

I want one that uses a pnumatic ram to toss salesmen off my property, after they ignore the 8"x11" NO SOLICITING THIS MEANS YOU sign.

One that can sense when I bring a visitor home and clean up real quick.

One that uses a deep booming voice to say "THAT SCUM IS SCAMMING YOU!" when the A/C repairman says my unit will colapse into dust in four seconds, posining my family and pets.

I want one that will project a hologram of Satan when Jehova's witnesses come around, or one of an avenging angel when satanist's come knocking.

One that will order booze when I'm low and send the bill to M$.

One with a radio control unit built in to mow the lawn.

Scans my e-mail and ZOT's spammers, then delete the e-mail.

Seriously, I don't know the relm of possibility with neural-networks. Haven't looked at 'em, don't know what's possible and what isn't. Guess I need to start looking around for info....