Artificial Intelligence can be made to do many things, and that is the magic of making computers smarter beyond its original programming.
But to make a system smart using AI and machine learning algorithms, it needs to endure the process of training. Before training, the AI knows nothing. But as it learns, it progresses and becomes better in doing a given task. The better way to describe this, is by showing it visually.
Software developer Keiwan Donyagard has developed an interactive AI demonstration in the form of an evolution simulation. Donyagard simply calls it 'Evolution', and the name clearly suggests what it means.
First of all, users are presented a blank page where they can experiment by creating their own creature by building rudimentary stick figures. Or, they can just select one object/creature from the list.
The dots represent the joints, the lines as the bones, and connecting them are the muscles.
Once assembled, users can simply "evolve" it.
The program then starts by unleashing several of the users stick-figure AI (population), trying to accomplish tasks, including running, jumping, obstacle jump and climbing.
While the evolution part is handled by the neural network that represents the creature's brain, what users do is creating the basic of the creature's body. Using stick figures as a representation of the creature, users can clearly see how the muscles expand and contract to give the bone movements with the joints as axis.
The first generations of the AI should be clumsy, dumb and incompetent. This is normal as the neural network has still to learn how to use its given body to accomplish the given task.
After a specific amount of time, the stick-figure AIs will die. This can be set using 'seconds per generation' option. They are then replaced by a new generation, but using only the best selected members of the previous generation. As the process continues, the AI should progressively becomes better.
In version 2, the system features optional grid to help users in placing joints, a customizable neural network size, option to watch the creature one at a time while they simulate, option to use multiple batches, adjustable advanced simulation settings, and more.
Donyagard's 'Evolution' is not a game, but only a simulation. The AI has no objectives.
Even if the creature created there reaches 100 percent capability (fitness level), 'Evolution' is just for excitement and joy, illustrating how AI improves by learning and failing.
It's like playing God, literally.