Described in these pages is an artificial intelligence, a learning algorithm, that is used to find patterns within a time series. The acquired time series is a measurable quantity, a collection of data points that vary over time. Collected from the wild it can represent almost anything that changes in time, such as: the count of sunspots, the number of hot dogs sold at a street corner per day, the weekly price of aluminum, etc. The time series to be analyzed is labeled as the output.
The output time series is generated at the back end of an unknown unviewable system: the black box. At the front end of the box enter the forces and variables that help produce the output time series, these are called the input and are considered unknowable and unmeasurable. Combining the input with the inner workings of the black box produces the collected output time series.
The algorithm searches for the hidden patterns within the output time series by following a set of steps that will try to mimic the unknown black box, thus, it tries to recreate the system that initially produced the output time series. Once an approximate model for the black box is found, the given time series (the output time series) is shifted into the future through the derived black box to generate a possible prediction of the system at the output.
The black box consists of a feed-forward neural network. It is trained using a back propagation algorithm, enveloped within a genetic algorithm. A random noise generator with various distributions is employed throughout the architecture of the network.
Below is a diagram of a single neural network. It consists of input neurons, shown as the five objects on the left, hidden neurons, the middle connected objects, and a single output neuron, the single object to the far right. The connecting lines represent possible pathways for information flow connecting each individual neuron.
Techniques from time series analysis and global optimization are used to help the algorithm arrive at a stable solution for the best neural network, which ultimately produces the resulting prediction. For example, applications in time series analysis suggest that the information about the unknown black box model and its input signals are embedded in the output signal. The algorithm's analysis of the output signal is used to solve for the unknown properties of the black box and input signal.
Problems of parroting are solved in a unique manner, allowing the neural nets to settle upon useful solutions.
The A.I. is capable of analyzing time-dependent
data, such as the price of raw materials, the price of stocks, or
the magnitude of any pertinent sequential data, for the purposes
of forecasting future patterns.
Below
are further descriptions of the inner workings of this A.I., along with some examples of its past performance. |