The last one will blow your mind…

Let’s be honest for a second and leave Terminator and HAL where they are, artificially intelligent systems are not going to dominate the world tomorrow. It does not mean however that AI systems are not already present around us, working for and with us at different levels. You probably have some examples in mind, whether it is AlphaGo, facial recognition or automatic handwriting recognition, but there might be some other domains that you don’t expect. So let’s go and get a round-up some of the domains in which AI systems are already implemented and working. This article is obviously not exhaustive about AI applications and many others exist.

1. Recommendation systems

You all know this little box in various e-commerce websites : “Our picks for you” or “Things that might interest you”. These sections give you recommendations of things you might want to buy depending on what you have bought previously, but also depending on what others who have bought the same thing as you also bought. It works for online shops of course, but the same systems are available for suggesting you movies or series to watch, music, or books you might enjoy.

Usually, three types of algorithms are considered for this task:

  1. Item-based recommendation: A profile is created for each item containing its characteristics (genre, winner of a prize…). The algorithm then finds the most similar items to recommend to you.
  2. Collaborative filtering: The algorithm looks for users who liked the same things as you and recommend what the other users also liked that you do not know.
  3. Hybrid approaches: Mix of the two previous ones1. For instance, the item-based collaborative filtering algorithm2 uses rating from users similar to you and selects the items that are “the most liked” among similar users.

For more information on recommender systems and a broader overview of existing techniques, I refer you to this article.

2. Games

Here is a second quite well-known application of AI! Whether it is video games or board games, a lot of AI systems have been developed to play games and be worthy opponents to human players. The range of techniques used for AI in games is so wide that it would require a series of articles just for itself. But in a (very) general way, AI in games is about planning. The artificial agent must decide on the next best action(s) to perform depending on the state of the game.

Recently, Reinforcement Learning has got a lot of attention for playing video games. Maybe you’ve already seen this type of video that illustrate very well Reinforcement Learning in a game such as Mario:

The idea is basically to try different possible actions and see how it goes. In the example of the video, it means that if Mario dies, then we consider that the succession of actions was not good. If Mario finished the level with the maximum points, then we consider the sequence was excellent.

It is also reinforcement learning which is at the base of AlphaGo, the program that defeated Lee Sedol in 20163, as well as AlphaStar, the program that defeated two Starcraft II pro players in 201945

Reinforcement learning is not the only way AI is present in games. As a matter of fact, it’s not even the most common at all despite getting all the spotlight recently. AI in commercial games is used in (roughly) two different planning categories: decision-making and pathfinding.

Decision making aims usually at controlling NPCs (Non-Player Characters) and give the illusion that they are pro-active, meaning that they are taking decisions themselves. One of the most common technique for this purpose in commercial video games is still good old Decision Trees.

A Decision Tree is a model in which each “node” of the tree represent a decision point (e.g. should I put my token in cell 1, 2 or 3?) and the edges between nodes represent the possible outcomes. For instance, this is an example of a decision tree for the Tic-Tac-Toe game:

First two layers of the decision tree for the Tic-Tac-Toe game. From Wikipedia, user:Gdr, CC BY-SA 3.0

By creating and combining decision trees at different levels of abstraction6, we can create characters that look like they are (more or less) intelligent.

Pathfinding is another AI problem in video games: how does my character goes from where it currently stands to where I clicked in the most efficient way? Once again, most commercial games nowadays are using “Good Old Fashion AI” (GOFAI) techniques by modeling the environment as a grid. Each cell is associated with a weight which depends on the environment7. Algorithms such as A* and its extensions are then applied to the grid 8

For more information about AI in video games, the Wikipedia article on this topic is quite well done.

3. Transportation systems

If I tell you AI and transportation systems, I have no doubt that you will think first about autonomous cars. But even before driverless cars were becoming a reality, Artificial Intelligence systems have been widely used in the domain of transportation systems. First of all, did you ever realize that your good old GPS is an excellent example of AI Systems? The algorithm used to find the best route in your GPS is probably a variant of A*, that we already mentioned. And of course, modern GPSs make good use of Speech Synthesis to give you directions.

But AI systems are also present in situations that you might not expect. For instance, they have been successfully used to reduce traffic jams by monitoring traffic lights and activating them optimally9. Other systems are installed in locomotives to reduce energy consumption and failure rates. Multi-Agent Systems and Machine Learning are usually the star techniques for this type of application.

AI systems in cars already exist, not to render the car fully autonomous but to be of assistance to the driver. For instance, collision avoidance systems detects obstacles and ensure that the car brakes on time, parking assistance systems help you fit in this very small but only spot. You can see this article for a good overview of what is currently done. Machine Learning, but also Sensor Fusion are the two big actors of this domain.

4. Logistics and Warehouse Management

Logistics and warehouses management are two very complex problems. They depends on many parameters difficult to model, some of them even time-dependent. For instance, transporting goods from where they are to where they should be create many challenges that AI systems can help to overcome. You might have already seen the Amazon warehouse’s robotic shelves, which make stacking and piling more efficient:

The Amazon robotic shelves. From MIT Technology Review.

Coordination and cooperation are key aspects of this type of application, and both are still very much studied. Motion Planning, Task Planning algorithms, as well as Multi-Agent Systems, are among the keywords to look for to understand these systems.

Another big domain for AI in logistic systems is called Anticipatory Logistic. Habitually based on Machine Learning techniques, it allows companies to predict spikes and drops in the demand and adapt their stock accordingly.

This Techzone article is very interesting on the topic. As well as the MIT Technology Review article on Amazon’s mobile robots, whose tasks go well beyond transporting shelves.

5. Design and Architecture

I bet many of you didn’t expect this one. As far as I am concerned, I didn’t see it coming the first time someone mentioned it to me. How on Earth could AI and design collide?

It is still relatively new both in research and applications, but it occurs that AI is a wonderful tool for Evidence-Based Design. Similar to Evidence-based Medicine which is medicine based on facts, Evidence-based design is based on empirical and quantifiable measures. AI techniques are used to help designers and architects to make rational and motivated decisions. Techniques based on Ontologies and Spatio-Temporal Reasoning are for instance of use to plan the building shape and convert the interior space.

Even after the building is built and in use, other techniques, especially using eye-tracking and wearable sensors are used to analyze how people actually use the building and space10. Data gathered this way offers new lights and insights for future designs and architectures.

Eye-tracking experiment showing where the subject is focusing. From (Bhatt2016)11

Conclusion

There are so many other applications that I didn’t mention in this article. From medicine and image-reading systems that help doctors, conversational agents used for fun or learning, virtual assistants such as Alexa or Mycroft… AI is already around us and will be even more in the upcoming years. Many projects are now being tests and will become products and used. AI is already changing our environment and lives, and it will keep going.

References

Here is the list of the sources used while writing this article and interesting to look at:

Recommendation systems

  • Melville P., Sindhwani V. (2017) Recommender Systems. In: Sammut C., Webb G.I. (eds) Encyclopedia of Machine Learning and Data Mining. Springer, Boston, MA
  • Sarwar, B. M., Karypis, G., Konstan, J. A., & Riedl, J. (2001). Item-based collaborative filtering recommendation algorithms. Www, 1, 285-295.

Games

Transportation systems

Logistics and Warehouse Management

Design and Architecture

  1. Thanks Captain Obvious
  2. A bit old but still gold
  3. https://en.wikipedia.org/wiki/AlphaGo_versus_Lee_Sedol
  4. https://deepmind.com/blog/alphastar-mastering-real-time-strategy-game-starcraft-ii/
  5. Despite some very interesting and impressive results of AlphaStar, I personally had some remarks and concerns about the way the games have been showcased and advertized. If you would be interested, you can find the thread I wrote on Twitter at the time. Some of the points have been made clearer since (such as the “focusing” part which is not relevant anymore) but most of the arguments remain valid in my opinion.
  6. From the global high-level strategy to the very detailed sentence selection in a dialogue option
  7. This allows the designer to model how difficult is the terrain, whether enemies are present…
  8. For information, the A* algorithm has been first published in 1968 and is an extension of the Dijkstra’s algorithm, itself published in 1959.
  9. You can look for instance at the Surtrac System for an example of such monitoring.
  10. As opposed to the expected usage.
  11. Bhatt, M., Suchan, J., Schultz, C., Kondyli, V., & Goyal, S. (2016, March). Artificial intelligence for predictive and evidence based architecture design. In Thirtieth AAAI Conference on Artificial Intelligence.
0 replies

Leave a Reply

Want to join the discussion?
Feel free to contribute!

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.