EV-BaseGame

Evolutionary algorithms to play basic games

Experimental Examples

Survival Rate

Gen 1:

Gen 10:

Gen 25:

Gen 100:

Experimental Examples 2

Survival Rate

Gen 1:

Gen 10:

Gen 25:

Gen 100:

Sample Mind-Map:

Final Generation's Map:

As we can see that when y is large, the agent is at the bottom of the map, which makes sense why the move_north neuron is activate. On the other hand, when x is too large, which means agent is on the right hand side, it activates the move_west neuron which pushes it towards the left side. The move_west neuron also gets deactivated with large values of y which means it potentially prefers to cover the y distance first.

Interesting Maps:

Taking a look at this we can see that when a pixel appears directly to the west(02) of the agent it reduces it's response-impulse towards move_west implying that it offers the agent to move north. This builds to the concept that upon blockage the agent has learnt to overcome it.

Input Representations

py { "prev_direction" : options ("north", "south", "west", "east") --> default (Any of the four) "moving" : options (0 - `previously_at_rest`, 1 - `previously_moving`) --> default (0) "x" : range (0, WIDTH) "y" : range (0, HEIGHT) "surround_space_pixel_00" --> options (0 - `empty space`, 1 - `map_border`, 2 - `another_bot`) --> default (0) ... "surround_space_pixel_13" --> options (0 - `empty space`, 1 - `map_border`, 2 - `another_bot`) --> default (0) ... "surround_space_pixel_44" --> options (0 - `empty space`, 1 - `map_border`, 2 - `another_bot`) --> default (0) }

The model basically receives a total of 2 (self-based) + 2 (global) + 24 (surroundings) --> 28 inputs.

Defining agent outputs:

py { "next_move" : options ("rest", "north", "south", "east", "west") # "forward", "left", "right", "backward") }

The output must be a single probability distributions allowing the individual to make the next move. --> SOFTMAX (5 neurons in the output wing)

Inner Neurons:

py // TO BE DEFINED

Understanding the genomics:

Input context - 28 input_variables and 5 output_variables. Reconstituting the individual inputs to be representable.

28 input_variables --> closest 2 power is 32 --> basically val=(00000) to restrict space, to the 28 possible inputs, we mod val with 28 (val%28). Can extend by 4 new inputs.

5 output_variables --> closest 2 power is 8 --> basically out=(000) to restrict space, to the 5 possible inputs, we mod val with 5 (out%5). Can extend by 3 new outputs.

Now we want float values to represent the passing weights between the inputs/outputs. Lets begin understanding the sensitivity of the softmax function, we can simply extend the sigmoid function's sensitive to approximate softmax, as sigmoid is simply limited softmax.

Looking at the graph, it should be substantial to consider upto 3 decimal places ._ _ _ at the same time, a bar between [-4, 4] should be sufficient. Therefore we need to find a numeric range constituting an integer space [0, 8000] ~ [0, 8192) = [0, 2^13). It will then be normalized as (value-4096)/1000.

| 0 | 00000 | 000 | 0-000-000-000-000 | basically a total of (1+5+3+13)=22 stream of binaries.

Representation:

Although this isn't an essential representations are essential for visualizations. So lets discuss representations, now human DNA is represented by:

md 1. adenine (A) - 00 2. guanine (G) - 01 3. cytosine (C) - 10 4. thymine (T) - 11

Basically it can encode two bins (00). So, we can represent our entire stream by a stream of 11 (A/G/C/T). Colour representation - pixel colours are an interesting aspect of visulaization so, what we do is represent them by RGB=[0,256), therefore we can extend our stream of 22 bins by 2 bins, making it 24 bins. Divide it into 8-8-8 bins allowing us to compute RGB values.