We evolve agents with twelve (12) binary node-networks (neurons), where the assigned task is to navigate through a 2 dimentional maze using appropriate navigation rules (to be learnt) and make least possible mistakes. At regular intervals, we measure the complexity of the evolved connectome, using some of the well-known complexity measures. The connectome is completely specified in the genome of the agent: genome decides how these neurons are connected with each other, which in turn decides how the agent would interact with the environment, here a 2 dimensional maze. The agent can sense its environment through 5 sensors: out of the total 12 nodes, three are front-collision sensors or retina, two lateral collision sensors. It can move laterally or ahead using two motion actuators. At each maze-door, the agent is informed about the lateral position of the next door via a binary door-information center (a red arrow = doorbit is set to one; and otherwise zero). The agent has to evolve the correct (internal) circuitry to remember this information and use whenever needed. The agent can use the remaining four nodes for this purpose.Note, the agent has to learn the "rules" of the game and not one particular strategy. So, we redesign the maze every 100th generation.
Below are movies of three incidences, where an agent (of course, different every time) is trying to navigate through a maze, per its evolved-so-far degree of intelligence.The lower half of the screen shows first the brain-connectivity and later the firing-like activity.
| |
|
|
||
|---|---|---|---|---|
|
|
|
|
In the middle is a result of an evolution through 60k generations. This particular agent from the population has learnt to correctly save the door-information in some memory-like unit and use it later. This can be checked via running a knockout analysis on the connectome. It has a fitness of 92%.
The rightmost is a carefully designed "Einstein" by the gods of the game (did someone hear saying, "we"?). The design seems to follow, Braitenberg's principle of "downhill-invention": see how tightly (and compactly) the wiring is achieved in just two (Markov) units. The connectome is not the most optimal when measured in terms of information-theoretically defined complexity measures, it seems.