Hopp til hovedinnhold

Juan Diego Cardenas Cartagena: Helping machines to make safe actions

Artikkelen er mer enn ett år gammel, og kan inneholde utdatert informasjon.

Juan Diego Cardenas Cartagena

(Denne presentasjonen holdes på engelsk)

Let me tell you the story about the time when I learnt how to cook rice. It is a simple recipe of two cups of water per one cup of rice, adding one tablespoon of salt on low heat for 20-ish minutes. The first time, I almost burnt the pot because the heat was too high. The second time, I got a fantastic rice soup because I added more water than necessary. The third time, the rice was too salty. It took me at least two months of trial and error to learn how to make a tasty bowl of rice. 

From this story, notice the trial and error approach for learning how to perform a new task. We use this approach to determine the required steps for reaching a goal and discover actions that may end up in troublesome results. What if, instead of cooking rice, that task is something that requires more care, like driving a car or managing the Kristiansand’s water distribution network. An error while learning those new skills is unacceptable. On the one hand, we may have a car accident with injured people; on the other hand, all Kristiansand may run out of water for days. To reduce those unwanted outcomes while learning a new task, we can watch tutorials, talk to experts, or take a course before doing any action. 

Even though machines may also learn new skills via trial and error, it is challenging to interact with them since their language is based on numbers. For this purpose, I am investigating techniques for machine learning without errors by formulating equations and logical steps that describe the risk behind taking some actions. At the end of the day, I am looking for a methodology to teach machines how to perform a new task while keeping safe behavior without undesired events.