NB: This project is built on the p5.js framework. See Technology below for links.
To experiment live, see: Micro Automata on p5.js
This project aims to extend upon the ideas of Conway's Game of Life by creating a simulation much closer to true microbial life. The simulation allows for 4 cell types: food (green), bacteria (blue), and viral phages that attack the bacteria (red). I implemented a preliminary version of this concept in 2012 on Android; see: MicroLife. This implementation is a from-the-ground-up-port to JS as to make it available in the browser. It also incorporates several new concepts:
- Neural Networks - Bacteria operate on a neural net that takes inputs (avalable food, current health, if it's infected by a phage, how many neighboring bacteria there are, and how many of those neighbors are infected by phages) and generates behavioral outputs (eating rate, cost of reproduction and health of offspring {costlier reproduction produces healthier offspring}, and if the bacteria should kill neighbors to reproduce).
- The neural nets are randomly structured with between 3 and 7 layers, each hidden layer having between 3 and 10 neurons.
- All produced neural networks are fully connected (FCNN) and do not have any convolutions nor recurance.
- The neural network was implemented using object oriented concepts (rather than a matrix approach) purposefully, trading performance for ease of understanding.
- Mutation - The neural networks have some chance of mutation which will effect the weights, biases, and mutation chance for the individual cell. This changes the behavior of the bacterial cell in random ways: some positive, some detrimental.
- Lineage Tracking - With each mutation a new bacterial lineage is created. This can be visualized by hovering over "Bacteria Species" which highlights all mutated cells with yellow, hovering over the ID of the bacteria which will color all original lineage with green and all mutations of that originating lineage in red, by mousing over the grid and inspecting the visualization of the neural net, or by inspecting the plotly-generated lineage chart. Examples are demonstrated below.
- Seasons - The rate at which the food grows varies "seasonally," peaking mid-summer and bottoming-out mid-winter.
The growing field defaults to a pure green, denoting abundant food. The colors of each cell in the field correspond to the levels of each corresponding state.
- Green - Food
- Blue - Bacteria
- Yellow - Phage
- Black - Wall
Walls cannot be traversed by bacteria nor phage, however a diagional gap can potentially leak if a bacteria or phage reproduces through the gap.
Field Full of Food
Field With Bacteria
Field With Bacteria and Phage
Field With Walls Encapsulating Different Species
The simulation will oscilate through sinoidal seasons centered around the Food: Growth Rate slider. Summer has the most abundant food while winter has the least.
Drawing types (food, bacteria, phage, wall, none) can be changed by mouse-scrolling.
There are several sliders available to adjust some of the environmental parameters:
- Food: Growth Rate - How quickly food on a cell grows (even if there is bacteria present). Seasons will center around this value. The more food, the faster bacteria can grow and longer they can persist.
- Phage: Voracity - How quickly phages will consume resources from their bacterial host. The more voracious the more quickly they will reproduce, but alsot the more quickly they will kill their host.
- Phage: Chance of Spread - When a phage spreads how likely it is to infect each neighboring cell.
- Frame Rate: How quickly the simulation runs.
- Cell Size (Resets) - The size of each cell on the screen. The smaller the size, the more cells. More cells typically looks better but runs slower. Changing this value resets the simulation.
This section shows how many mutations have occurred and metrics for each bacterial species, identified by ID. For each specias a count of how many bacteria are currently alive is shown. Mousing over Bacteria Species: will highlight all mutated cells in yellow. Mousing over an individual species will highlight it in green, will highlight any evolutionary relatives in red, and will show the bacterial neural net.
This visualization of the neural nets shows input neurons on the left and the output layer on the right. Weights (edges) and biases (nodes) are all on the scale of -1..1, visualized as -1==black and 1==almost-white.
As this creates output values in a quasi-Gaussian distribution centered around 0. The output values are expected to be in the range 0..1 with a non-trivial probability of reaching boundary values. To accomplish this the range is projected to the range -4.5..5.5 (a range of 10 centered around 0.5) and trimmed to 0..1. See: NNNeuron.outputToZeroOneRange.
Seasons introduce evolutionary pressure by changing the amount of food available to be consumed by the bacteria. Often bacterial species will simply die out during this period. Some go into hibernation. Some exhibit a ravenous sparse behavior, where they move quickly but don't reproduce in all directions as to consume as much food as possible while leaving a trail of wasteland behind them. In some circumstances a mutation introduces a variant in behavior that creates one of these behaviors, making the bacterial species more resilient than it previously was.
Example 1:
This simulation starts with three bacterial species (bacteria that never exceed a populatoin of 1 are not represented): 2, 4, and 8. Species 4 doesn't last long, but Species 2 and Species 8 begin to grow. Eventually, Species 2 and Species 8 run into one another and Species 8 quickly overtakes Species 2, leaving Species 8 as the sole remaining species.
Species 8 then grows to fill the entire field through Summer and into Fall. The population dips in Fall/Winter, but rebounds in Spring. However, in Spring a mutation of Species 8 makes an appearance: Sub-Species 8->37. While there is space to grow Sub-Species 8->37 grows, but once the field is again saturated Species 8 out-competes Sub-Species 8->37 and if the time of abundance continued it seems likely Sub-Species 8->37 would have gone extinct. However, come Fall/Winter Species 8 dies back leaving space for Sub-Species 8->37 to grow.
Species 8, with its lower population isn't able to pressure Sub-Species 8->37 during the third Summer/Fall and the competition is a stalemate. However, come the third winter, Sub-Species 8->37 is able to out-compete in the time of scarsity and pushes Species 8 to extinction.
Dozens of other sub-species and sub-sub-species can be seen that don't out-compete the dominant species, never gaining a foothold.
Example 2:
This example is a more complicated example, showing a longer timeframe with a more complicated lineage. Note, this all originated from an oringal seed of bacteria. All sub-species are descendents of a single bacteria, Species 15. A phage was introduced at Tick 4000 and another at Tick 6000. By the end of the simulation a 5th generation descendent of Species 15, Sub-Sub-Sub-Sub-Species 15->84->190->504->578 was the only remaining variant.
Sub-Sub-Sub-Sub-Species 15->84->190->504->578 and the phage
The introduction of a phage often places pressure on systems to exhibit new behavior. For example, a common pattern when pressured by a phage is to favor mutations that produce less-healthy offspring. Counterintuitively, this helps prevent the spread of a voracious phage as the weaker hosts die off before the phage can reproduce.
- Control the neural network topology via configuration.
- Add the ability to follow a bacteria (see its activation inputs and outputs).
- Neural Nets for the phages.
- Neural Nets for food.
- Ability to recognize relatives.
- Introduce a headless "tournament-style" mode.
- Develop characterizations framework (greedy, nice, neighborly, agressive).
- Develop a loss function and introduce a gradient-descent optimized evolution. Possibly a random bacterial variant that has this turned on.
- Possibly introduce NN structural variants: RNN & CNN.
- Add more inputs/outputs (e.g. age or self-destruct).
- Add vision and/or directional inputs and outputs.