In order to realize the difference between the mental abilities of a 7 years old kid and artificial intelligence, you should check out the famous video game Minecraft. Surprisingly, a young child can easily learn to find the rare diamonds by just watching a 10-minute YouTube guide. A recent computing competition organized by MineRL shows that researchers may try to reduce the gap between child’s brain and AI-powered machines. This experiment may further reduce the training efficiency of AIs.
The competitors may require almost 4 days to train AIs to find the diamond, and it can be done by following more than 8 million steps. The efforts and time taken are much more than a child may require learning the same thing. This contest is basically organized to spread awareness about a new approach: imitation learning. The idea is same as that of reinforcement learning where programs keep on doing millions of random things to learn best information. This technique has been already used to train robotic arms, to generate recommendations for various Netflix users and in the gaming world as well. However, it is important to mention that these tasks require much computing power and time as well. In order to create algorithms to drive a car or to execute many other activities in the games, the developers may need to put thousands of computers on parallel processing. It is not possible for all governments to afford such a huge cost for heavy resource utilization.
On the other side, the concept of imitation learning is given to power up the learning process by simply mimicking the activities of other AI algorithms and human beings. MineRL coding event also encourages participants to teach AI to play a game like humans. A Ph.D. candidate from Carnegie Mellon University, Pittsburgh, William Guss, recently said that Reinforcement Learning techniques do not have potential to be a part of such competitions. He also added that AI could be used to train a system to chop trees within the limit of 8 million steps. And the fact is that it is just a small part of the activity required to train to find diamonds in the Minecraft.
There is no doubt to say that imitation learning has the power to train the system about the environment. Guss and some of his colleagues also said that this contest sponsored by Microsoft and Carnegie Mellon might inspire coders to work on the extreme abilities of imitation learning. Such research activities can be more useful to develop new AI systems that can have more efficient interactions with human beings in different routine life situations. Furthermore, it can also boost the abilities to navigate environments that are loaded with complexity and uncertainty. A research scientist at Google DeepMind, London, Oriol Vinyals, also said that Imitation Learning could be observed as the core of several developments in the field of artificial intelligence. It can help machines to learn tasks quickly without working from scratch every time.
Gaming by example:
Researchers involved in this competition revealed that Minecraft is a great example of virtual training. The players reflect many intelligent behaviors while playing this game. While playing in the survival mode, they need to protect themselves from monsters while collecting materials to build structures. The new-age players may need to understand the physics hidden behind this game and collect ideas about how materials can be transformed into valuable tools or resources.
In order to collect some valuable training data to execute competition, organizers at MineRL created a public server of Minecraft and asked some people to take the challenge of executing few specific tasks; for example, crafting different tools. During this activity, they captured more than 60 million examples of various actions taken by players during different situations. These recordings are further going to help the scientific community to go deeper into imitation-learning research.
The idea is to use imitation learning as bootstrap training. As a result, AI may not require so much time to start activities from scratch; rather, they will utilize the knowledge that is already developed by human beings. Although the concept of reinforcement learning has contributed so many research studies around the world. But in the coming days, imitation-learning would be the center of attraction for Minecraft server hosting. The biggest benefit of using Imitation learning is that it reduces the efforts carried out on trial-and-error approach.
Let us consider the example of self-driving cars. In order to train those cars through the reinforcement learning approach may require millions of trials to achieve safe driving results. But in actual, the simulation-based data cannot train a system to make appropriate decisions in the critical crash situations of real world. On the other side, allowing cars to crash on road to collect data regarding the event may be a dangerous activity. In such situations, imitation learning can work more efficiently than that of the AI.
There is no doubt to say that imitation learning also has some limitations. The biggest one is that this technique is a little biased towards the conditions demonstrated in the input learning examples. In case if AI systems are trained with such data, they may tend to become more inflexible. The chances are that if AI makes some mistakes, it may diverge from required flow of algorithm and create a different setting as compared to the demonstrated environment.
However, even after such limitations, the new age researchers find great potential in this technique. The best part about imitation learning over reinforcement learning is that it allows you to get a clear demonstration of the success of any task.
In order to handle the diamond situation, the players at MineRL may have to generate a multi-step sequence. They may require hardware with much higher power to execute heavy programs. Researchers believe that the complexity of problems involved in the MineRL contest may encourage the progress of imitation learning. The findings will be more useful for real world applications in the coming years.