Egg-Bot Triumphs!
Mr.Eggy surviving the zombie apocalypse.
Waking up this morning, I couldn’t wait to crack open the results from the Egg-Bot’s overnight training. And boy, did it deliver—egg-straordinarily well! Mr.Eggy has turned out to be a real yolk star. His legs are dodging zombies like a butterfly, and the turret is stinging like a bee. Mr.Eggy just shattered records with a marathon game lasting over an hour! The Egg-Bot absolutely nailed the behavior we were aiming for—watching it in action is like seeing a perfectly cooked omelet flip in the pan.
We took the complexity of training and scrambled it by splitting the tasks into separate neural networks. The outcome? Still egg-cellent! Besides the ray sensors picking up on zombies, each network got only one extra observation: its own position. The legs had a singular incentive—to stay alive—while the turret had two: stay alive and shoot zombies. With these simple straightforward inputs, the Egg-Bot learned our desired behaviors, proving that our hypothesis wasn’t just a bunch of half-baked ideas.
This success has me all scrambled with excitement about future possibilities. I’m keen to try out the turret or legs independently. For instance, what if we mounted the turret on a pillar instead of the legs? It could add a whole new layer of modularity.
And since these networks are working independently, there’s a cracking chance we could combine them into a super team. I’m considering giving the MA-POCA algorithm a spin instead of the PPO, allowing the networks to unite under one team ID with a shared reward system.
This fresh approach might just whisk away my old methods of network training. Think about it: training separate networks on simpler tasks and then stacking them together like LEGO bricks—connecting outputs to inputs or building them up to form a more intricate system. It’s like creating a masterpiece with tiny, precise pieces.
Stay tuned as we continue to refine our ML-Agents! This journey is just getting started, and we’re just getting started in the world of AI—egg-spect more egg-citing developments soon!
