Interactive demonstration of Neural Cellular Maze Solver. This cellular automaton is trained to output the shortest path between the two endpoints. You can interactively edit the maze input by clicking or tapping with selected maze cell types (Wall, Road, Endpoint). The state of each cell is stochastically updated depending on the state of each cell and the four-surrounding cells.
Swarm intelligence is a system in which each of its components only has a simple function and these components interact only with their surroundings to produce complex results that are unimaginable from their individual functions. For example, in the study of mouldy computers, they found that slime mold solves the shortest path problem. When we place slime in the maze and feed on endpoints of the maze, only slime mold on the shortest path remains. Although each slime mold cell is not intelligent enough to calculate the shortest path, slime mold emergently solves the shortest path problem through the interaction between the cells and the environment. Like this example, agents with simple functions have emergent functions as a group through interaction. So we can consider createing artificial swarm intelligence to solve other problems. Boid is famous for the first artificial swarm intelligence to simulate a flock of birds. Emergent results are observed that cannot be discovered by only knowing how the agents interact with each other. Therefore, it is a challenge to design the functionality of agents to achieve the desired result.
Growing Neural Cellular Automata showed that cellular automata based on simple rules can self-organize a variety of desired complex structures. In this model, the next state of a cell is determined only from the state of the surrounding cells, and the same rules are used for all cells. Since the cell update rules are differentiable, the process of self-organization can be optimized by end-to-end learning. In the Self-classifying MNIST Digits article, they showed that cellular automata are capable of recognizing numerical images. They assign each pixel as each cell on the grid and determine what numbers are written on the pixel as a whole. In this case, cells can only obtain information directly from neighboring pixels, so it is necessary to obtain information by recursive interactions between cells to get the overall picture. In other words, they showed that the model is possible to learn how to integrate information on a global scale based only on local interaction.