(May need to click to play)

I'm simulating a collective of independent agents. Each is controlled by a set of neural networks that determine the agent's action. The networks train themselves based on positive and negative action outcomes. Currently they eat, swarm, and reproduce. Above you can see them moving as a collective to consume food (green). Their population (plotted below) fluctuates pretty steadily and you can see the group expand and contract based on environmental stimuli. It's pretty processor-intensive, so I've had to set the environment to support only a few hundred organisms.



The algorithms are still pretty basic, development goals include:

• More interesting environment features.
• Physical traits beyond the decision networks and lifespan.
• A genetic algorithm for generating new agents.
• Faster pseudo-random and pseud0-gaussian number generators, with minimal flops.
• Heritage records to determine which founding organisms were most successful.
• Multiple antagonizing colonies.
• UI to allow realtime environment manipulation.Leave me your thoughts on algorithms and features.

Comments

Related / internal

Related / external