AI Hack Night: Introduction to artificial intelligence

February 24, 2016

Artificial Intelligence (AI) is a regular topic of discussion at Arq Group. Many of us are interested in the direction of AI research and love to analyse the breakthroughs that are frequently being made in the field. So when we got the chance to organise another of Arq Group’s hack nights, it simply made sense for a group of us to get together and try to teach ourselves how to apply some of the techniques in practice.

 

Three AI challenges

We set out to achieve three simple goals:

  1. Complete some of the introductory HackerRank AI exercises
  2. Triumph over a mid-level RoboCode bot
  3. Train the MetaMind image-recognition bot to recognise a new class of image

The reward for completing these challenges? Eternal glory!

And chocolate, lots of chocolate.

 

Hackerrank

Wanting to start off simple, we began with attempting the BotClean challenge on HackerRank. The challenge is essentially a version of the Travelling Salesman problem, attempting to visit a number of locations in the fewest moves possible. This refreshed our knowledge of search algorithms and got us in the right frame of mind to move onto more difficult challenges.

 

Robocode

After triumphing over the beginner exercise we transitioned to experimenting with battle-bots in the RoboCode playground. Most of us quickly created bots capable of beating entry-level examples, but more complicated bots proved difficult. Implementing an effective bot can be a long trek down the path towards deep-learning algorithms, but we stopped at aiming algorithms and decision trees. Still, a few of us are keen to continue experimenting with this fun application of AI!

RoboCode

Metamind

Our final challenge was to experiment with the functionality exposed on the MetaMind website. We discussed some of the deep-learning techniques the MetaMind AI uses to solve image-recognition problems, then attempted to train it to recognise the difference between a strawberry and a raspberry. With a very small dataset (only about 10 images in both categories) we saw some fairly confident results, with the MetaMind AI being 79% confident that a new image of a raspberry was in fact a raspberry.  

MetaMind

Outcome

We all went in with limited practical experience and mostly theoretical knowledge about the state of AI research and techniques. Now that we’ve played with some hands-on examples we’re all more excited about the possibilities, and hopefully, understand a little more about the inner workings of basic AI. A big thanks to Arq Group and the organising crew for helping us explore this exciting tech!

AI Hack Night