Designing evolutionary self replication in physical CA with deep learning

[simsim314]

Deep learning with convolutional nets (cnns) is a tool which for a while now is demonstrating best performance in generalization of learned examples data efficiently and seamlessly. CNNs can provide a way to make cellular automata mechanisms that learn the evolution rules from movies i.e. showing a lot of movies as training data the net will take a cube of Dt NxN frames and predict the next frame cell by cell using taught evolutional CA rule. Thus we can expect to "mathematize" any movie into a CA with CNN.

Physics Simulation This approach can be applied directly to biological creatures movie footage that have self replication as well. Hopefully such trained CA will exhibit self replication and some amount of biological and physical traits. As a more "computational" approach I suggest a more "direct" design of self replicational physics that can be simulated with CA based on CNNs.

Breaking physics for self replication

Stage 1: Molecular physical training. In the image you can see the basic molecules design. It has two edges and it can attach by the yellow circle spontaneously. The red arrow shows the direction of movement. It moves and rotates by default just like regular physical object and the first batch of training movies should be done with simulating physical properties of this "molecules".

cratures0.png (126.56 KiB) Viewed 4838 times

Stage 2: Self replication training. Now we can imagine how self replication would occur. We should not think about the physics only imagination of self replicating process inside a "soup" of this molecules. We should produce none physical movies of self replicating proccess. This will be the second training dataset.

Notice that during the training the basic physical quantities like shape of the molecules, their amount and even momentum and angular momentum etc. are preserved, only the physics breaks sometimes to self replicate. Now we only left to make this experiment and hope for the "alternative" physics which will be guided by self replicating friendly mathematical creatures in CA and hopefully with some luck the creatures could become more and more complex, because the resources are limited. Some creatures would be able to use other creatures as building blocks and real "mechanical evolution" could emerge.