Will Machine Learning Ever Understand Skiing?
This is a difficult question, in my opinion. I’d say it depends on what you’re talking about.
Firstly, let’s divide the question into several parts: what are we attempting to understand? What becomes the proper / optimal skiing technique given physical characteristics, ski slopes, and atmospheric conditions? Is it possible to teach a robot to ski in general conditions?
Given the difficulty of the ski slopes and the skier’s experience, how should the skipass be limited?
I believe that with the right data, Machine Learning techniques can find the right patterns to manage those problems. Obviously, data collection is difficult, but it is not a nightmare.
Last year, I went skiing in Corvara, Alta Badia, Italy, and the Dolomiti Superski company was collecting geolocalization data with their skipass application.
Each of these issues can become addressed using a different method. I believe that the first question requires a high level of expertise in physics and engineering, as well as a wide range of techniques and domain knowledge.
Using data from previous alpine skiing winter Olympics, we can use unsupervised learning to determine the fastest and least risky possible path for the athletes who won based on their physical abilities.
The second question can be addressed using Reinforcement Learning techniques. I hope this is not a basic consideration, but whenever there is a problem where we do not know how to properly code a robot to do something, Reinforcement Learning and Deep Reinforcement Learning techniques come to mind.
However, given the number of changes that occur with different slopes and atmospheric conditions, teaching a robot to ski may be difficult without an enormous amount of interactions, which can be costly (reward shaping is impossible).
However, in my opinion, the best approach is to address it with a continuous reward system and to use hindsight experience replay to learn from unsuccessful trials as well, as they will constitute the vast majority. Regarding the third question, it is, in my opinion, the simplest to implement.
As I previously stated, companies collect data on skiers, and data on accidents is either publicly available or can still be collected via geolocation. A supervised system can become set up to prevent inexperienced skiers from approaching highly difficult ski slopes. Or they can become advised that the slope level is too high for them. And thus the insurance will not cover them in the event of an accident.
In conclusion, this can benefit both skiers by reducing accidents and insurance companies. All of these issues, and many more, can be addressed using Machine Learning techniques. However, what about truly understanding what skiing is? I believe that Machine Learning techniques are insufficient; perhaps conscious AI becomes required in that case, but based on what I have recently read, we are still a long way from that.