14 July 2018

How to effectively predict the future with AI

Modern artificial intelligence (AI) is very good at determining probabilities based on historical data. AI has reached a point where the average person can now understand forecasting potential of AI, particularly due to recent advances in AI over the past few years.

AI is now being used for everything from financial market prediction, to predicting what a daytime photo might look like at night, to predicting what a specific or a typical human being might sound like when reading out a text sentence.

Predicting the future is possibly the most intriguing aspect of artificial intelligence for me and it's an area that I feel could use a good systemic approach. The more data AI can get through the more accurate it's predictions, so it's logical to think that the more simplified the data the more data can be consumed.

It's important to note that AI predictions, at this stage at least, rely entirely on historical data to derive patterns. For example, if your son goes to soccer practice every Tuesday and Thursday, it's reasonable to for an AI to predict that he will score a goal either next Tuesday or next Thursday. If the data model does not include weather patterns, it can not predict for example, that a storm will cause  next week's practice to be cancelled and thereby reduce the goal scoring potential to zero.

Like all technology, Artificial Intelligence is not magic, it's applied science. But most importantly; AI deals with probabilities, not absolutes. That's important because, the finer details of AI prediction are almost always inaccurate but are close enough to build a very accurate broad picture.

So I've come up with what I feel is a simple way of storing event data (which is combined spacetime measurements) where only unique information is stored and the system relies heavily on extrapolation (something AI is particularly good at).

It's a wholly relative system so we use something fairly consistent as a centre point: the Galactic Centre. So take our star Sol, if we draw a line from the centre gravity of Sol to the Galactic Centre we can call that point on the Solar sphere the zero point. If we draw a line from the centre gravity of Sol to the centre of gravity of Earth we can then calculate the angle and rotation from the Solar zero point  and get a value for the direction Earth is relative to Sol. We then add a second component which is distance and now we know exactly where the Earth is in the solar system for that event.

We can then use our extensive understanding of the movements of celestial bodies to extrapolate the date and time from that location data (possibly relative to an agreed absolute date and time such as the start of the 21st century). We then use the line between Earth and Sol to set the zero point for Earth and we then calculate the direction and distance relative to that point for the hypothetical event that we're recording. Let's say that point corresponds to the entrance of Durban Harbour. We can then take it a step further and record the position of a ship relative to Durban Harbour.

So the data might be captured as:

Earth: direction [180.056° by 5.4376°] distance [149.6 gigameters]
Durban Harbour: direction [215.859° by 18.511°] distance [6.371 megameters]
Ship:  direction [90.0002° by 58.876°] distance [3.2 kilometers]

It would then need to be stored in a data efficient manner, maybe something like:

The bulk of the computation will be spent not on training the model but on converting event dates and locations into spacetime values. Once we have enough event data we can start training models on historical battles, or hurricanes, or car accidents, or maybe positive events like the time and place of Nobel prize winning discoveries or inventions, or the construction of historically significant buildings, etcetera.

The most valuable part of this system is that we can take any predictions produced and calculate some definite inaccuracies such as whether the Earth falls outside of it's orbit, or an event occurs too far away from a reasonable location such as on the Earth's surface. With this erroneous data, we can then re-train the model to make less erroneous predictions in future until we get to a point where we could very probably predict events of some significance to us, and possibly predict more detailed events such as probable car accident prone areas or areas of high probability of scientific advancements.