Neural networks may definitely be used to heat, cool, humidify and light a building to simultaneously minimize energy usage and maximize comfort. The potential benefits are reduction of fossil fuel usage, improved health and improved productivity.
Don't underestimate this difficulty of this problem, especially for large buildings with large windows that are delightful to work in but receive a lot of heat from sunlight during the day. Modern buildings must also bring in outside air for good health but not expend too much energy to heat and humidify that air. Heat and cold flows can be chaotic and unpredictable, leading to much discomfort. Simple logic and programmable thermostats are not good enough. In my opinion, a neural network is the best solution.
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.
Not necessarily. Any help for finding a solution can be used to assist the neural network possibly by input preprocessing or parameter search restriction. The power of the neural network is to solve the toughest part of the problem, say the effect of sunlight on heat flows or exterior doors on cold drafts (Light or door sensors are not necessary. These can be inferred from temperature data by the neural net.) Easily solvable parts of the problem may be incorporated into the design of the neural net. For example, the network might first learn how to reach a desired steady state from some other state. This fixes some parameters. Since heating and cooling behavior sometimes follow 24 hour and 365 day cycles, the network might try to solve for a multi-cyclical solution. Things like transient weather changes and people's behavior are harder.
Neural networks are not the only trainable networks out there. Linear networks with non-linear input stages work well as a first approximation. They can quickly find an exact (global minimum error) solution using a conjugate gradient descent or pseudoinverse algorithms. More advanced neural networks have been used to predict the stock market, interpret human speech and even drive cars at 60 miles per hour on a highway with a video signal as input and a steering wheel position as output. Someday, they may help to reduce automobile fatalities, or diagnose medical problems, or diagnose that a driver is driving drunk. I know of no more powerful algorithm for solving these hard computational problems.
Slashdot Mirror
Neural networks may definitely be used to heat, cool, humidify and light a building to simultaneously minimize energy usage and maximize comfort. The potential benefits are reduction of fossil fuel usage, improved health and improved productivity.
Don't underestimate this difficulty of this problem, especially for large buildings with large windows that are delightful to work in but receive a lot of heat from sunlight during the day. Modern buildings must also bring in outside air for good health but not expend too much energy to heat and humidify that air. Heat and cold flows can be chaotic and unpredictable, leading to much discomfort. Simple logic and programmable thermostats are not good enough. In my opinion, a neural network is the best solution.
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.
Not necessarily. Any help for finding a solution can be used to assist the neural network possibly by input preprocessing or parameter search restriction. The power of the neural network is to solve the toughest part of the problem, say the effect of sunlight on heat flows or exterior doors on cold drafts (Light or door sensors are not necessary. These can be inferred from temperature data by the neural net.) Easily solvable parts of the problem may be incorporated into the design of the neural net. For example, the network might first learn how to reach a desired steady state from some other state. This fixes some parameters. Since heating and cooling behavior sometimes follow 24 hour and 365 day cycles, the network might try to solve for a multi-cyclical solution. Things like transient weather changes and people's behavior are harder.
Neural networks are not the only trainable networks out there. Linear networks with non-linear input stages work well as a first approximation. They can quickly find an exact (global minimum error) solution using a conjugate gradient descent or pseudoinverse algorithms. More advanced neural networks have been used to predict the stock market, interpret human speech and even drive cars at 60 miles per hour on a highway with a video signal as input and a steering wheel position as output. Someday, they may help to reduce automobile fatalities, or diagnose medical problems, or diagnose that a driver is driving drunk. I know of no more powerful algorithm for solving these hard computational problems.