I recently learned how baby chicks follow their parent around, observing and copying, they do this to learn basic patterns to be able to adapt and survive. I wondered if the same could be applied in a game? The AI would consist of a neural net and AI bots would follow a human player around learning from their behaviours.

As a foundation I choose to use WebGL and the Legend of Zelda:Link to the past graphics. Besides being a great game, it also provides a rich environment.

I created a player control (Link) and Kiki the Monkey bots. The bots (Kiki) “watch” the player and try and follow him around, learning as the go by reacting to negative events. Each bot would have feelers, these are projected lines that look for contact with objects and walls. This contact feeds a negative input to the neural net, the harder the contact the more negative the weight, this naturally trains the net to avoid objects. Why code object avoidance when you can train it.

The net, when propagated for a result, returns based on the sensory inputs the best direction of travel. This also gives a great side action of if a player or other object get too close to a bot the bot moves away. As a group it appears like they have a central control routing them all in a complex manor.

Something I still would like to add is a reward system i.e. health pickups, this should had another layer of complexity to the perceived intelligence.

Steven

Update 24 June 2016

For future predictions I plan on adding a confidence level, so when you have such a small margin in a prediction result on a binary event it enables a more quantitative analysis.

Update 15 June 2016

Change in prediction. After reassessing the data now the final campaigns have pushed through, the vote is showing as a slight leave. During this campaign the data has been very turbulent and did show the gap narrow but still with a remain vote. It’s been interesting that a large percentage of a population can be undecided and swayed by the media.

Overview

I’ll detail at a macro level how my platform can automatically predict in real-time the outcome of the upcoming British EU exit vote (BREXIT). It has been previously used in a similar manner to predict US and British election results within less than 1% error rate. The components, advanced systems, and algorithms used to fully achieve this will be detailed in this paper at the business usage level to allow the reader to understand other areas of use for the platform.

How it’s done (in brief)

The main processes involved in predicating future outcomes are
     • Efficiently collect documents
     • Extract and normalize the data with the documents
     • Analyze and machine understand the normalized data
     • Cross-correlate the data and run advanced algorithms
     • Produce the present the prediction

This is all presented instantly and achieved with low latency. Each section listed has unique challenges and these will be outlined in the next section.

How it’s done (the science)

Efficiently collect document
Collecting documents efficiently provides many challenges. For the case of the BREXIT the documents are in the form of news articles collected from the internet. Internet web scrapers search the Internet and social media for news articles, and index these articles much in the same way google indexes the web. The scrapers run in parallel on multiple different cloud computing providers using low cost micro instances. These micro instances normalize and compress the data before sending to a central transactional server farm. This scheme provides global low cost redundancy whilst allowing for a linear cost over latency, i.e. the more micro instances, the more articles are indexed per second, but for more cost. Currently the scrapers running on 8 instances find and index 10 unique articles per second.

It’s currently indexing over 560,000 global news sources. For comparison Googles indexes ~25,000 for it’s news service. This is all in real-time with low latency, as having the latest breaking news is always critical.

Extract and normalize the data with the documents
First the meaningfully text from the web page is extracted and then run through Natural Language Processing (NLP) algorithms. This process is where the machine understands as close as possible to a human understanding of the entity types and context they relate to. The system auto discovers new entities along with their aliases, an example for this case would be “Brexit”, “British EU exit”, “GB EU exit” etc. These terms all mean the same thing and so the system should process them as a single entity. These entities and relationships are then placed into a custom built distributed graph database. The graph sits on top of Cassandra and also provides a linearly scalable data size and concurrent read/write access speed.
Analyze and machine understand the normalized data

Machine Learning (ML) and other custom algorithms are then employed to perform geospatial reckoning, sentiment analysis on five dimensions, and deep inter-entity relational links.

From this the sentiment for BREXIT can be quickly retrieved. It should be noted that for BREXIT a negative or positive sentiment is not an indicator of the overall medias intention of wanting an exit. An article could positively agree for either leaving or staying. This is where the graph DB is used to retrieve the complex entities relationship cluster link, along with time and sentiment.

For example, a typical cluster for a pro BREXIT could include:

BREXIT -> “immigration” positive

And for an anti BREXIT:

BREXIT -> “weak pound” negative

As can been seen from this, sentiment alone isn’t an indicator of a desire for an EU exit, the relationships are also needed to determine desire.

Cross-correlate the data and run advanced algorithms
This is the area where big data comes into play. Using statistical analysis from large data sets, biases and outlier data can be removed, thus keeping the ending results clean. Entity related trends are determined along with sentiment mood.
Further Spectral Regression is applied to predict the next 7 days based on historical patterns within the scope of the entity and its related cluster terms.

Produce and present the prediction
The results from the sentiment analysis of the clusters is then simply presented to the user as a percentage chance of binary event of Britain exiting or staying in the EU.

From correlating social media to news media, it was discovered that the news media sets the trends and opinion choices for the masses. The masses turn to social media to further express an opinion that they derived from the news media. Tracking unbiasedly the news media allows you to track the masses because of the scale of numbers. This is way the system works with such precise results, although simple in principle, it has been shown in the paper it’s complex in logistics.

Results

The data as of April 10 2016 shows a final result of 40% for Britain leaving the EU. Historically the data shows as always being a close vote, but in the run up to the final days the gap is lessening. From this consistent diminishing historical gap, the likelihood for last minute swings has a low probability, but still shows as potentially possible if a dramatic media campaign has still to occur.

In closing

The system discovers ~80,000 new entities per day. This number varies depending on new news topics. This is key to keeping the results relevant, tomorrows news won’t be the same as todays and over years’ large topic drifts would occur.

A good example is Apple computers. 20 years ago nobody would future associate them with phones and music and because the system can make these associations it can better sentiment analyze which in tern correlates to stock price because they are a very consumer driven company.

Another example is Donald Trump; in early 2015 someone asked if the system had a bug because it was showing Donald Trump in the top 8 for winning the US presidency. This was before he announced that he was running, so it looked ridiculous to see his name amongst the other candidates. I simply told the person who reported the bug that the system determined, from the data relationships that he may plan to run and if so has a statistical chance of winning. Further shock happened when the system increased his chances even through he was garnering so much negative press.

Steven

One of the new methods in machine learning is spectral regression (SR). SR works great for pattern recognition and so I built testbed in node.js to determine if it’s possible to predict future sentiment based on trained data.

I proceeded by taking 360 days of sentiment data for Apple and Google, and then split each into 3 sets. I cross validated the sets by using 2 for training and the last one for predicting, run for each of the 3 permutations. The 7 days ahead prediction values would be calculated on the prior 14 days of data. After running the 3 permutations I then sliced the 3 prediction windows together to create the predictions over 360 days.

The results are very encouraging, as can be seen from the data whilst the magnitude of the sentiment does sometimes have large over estimates the overall direction change does follow closely with reality.

• blue line is actual data, green line is predicted data

This does imply the ability to predict 7 day sentiment for keywords, and from my previous experiments with stock correlations, could allow you to determine future market indicators.

Steven

I’ve been collecting news articles and running natural language processing (NLP) and machine learning (ML) on them for a while.
I noticed that publicly traded companies heavily use PR to help increase their stock price. This lead me to thinking that maybe there’s a correlation between media sentiment and stock price.

I took the 90 day sentiment analysis for Apple (APPL) from my analytics engine and on this overlaid the closing price of the stock. Visually I could see a correlation.

• unfortunately I can’t overlay the stock closing data on sentiment as it breaks the data usage rights (can’t be reprinted in any form).

I then wrote a quick node.js program to run over all the NASDAQ listed stocks and correlate closing price with overall sentiment direction. From this I discovered most of the listed stocks don’t have enough media mentions to merit the volume of data needed to run the math. For the big stocks, there are definite correlations, most with a 1 day or same day lead time, a few like Apple/Google/Twitter that would give a 2 day buy/sell indicator.

My conclusion from this experiment is that large public companies that consumers have a passion for are heavily traded by the public and because of this media sentiment heavily affects the direction of stock as the news reaches the consumer investors.

Steven