Dr. Michael Williams

Congratitilations!

Congratulations, Michael!

Congratulations!

Congratulations.

Did they teach you the secret handshake right after the defense? I hear some universities do that but I had to wait till my diploma arrived.

Congratulations, Michael. It's quite a feeling, isn't it?

Congratulations, Dr. Williams!

Congratulations! I can't even imagine the relief you must feel.

Congratulations!

I'm glad you posted the slides. I read most of them, and your project was a lot more interesting than most. One thing that occurred to me while reading- you hypothesize that the Merit Hierarchy tribes does so poorly against Age Hierarchy because it is bad for the old animats to change their behavior because of commands from young, more successful animats. I am not a neural net expert, but that makes sense since the neural net is probably fairly static in the old animat, and a command that is very different from the neural net settings would tend to make the old animat act more chaotically, I would think.

It seems to me that there might be another reason for why the Merit Hierarchy does poorly. Since you create the new generations all at the same time, there are always multiple animats that can command the same sets of younger animats- i.e. they are co-commanders. In the Merit Hierarchy there is likely a top dog animat that can command everyone else, a second-in-command that can command everyone but the top dog, etc. Since your work has shown that the animats can and do carve out roles for themselves, those super-influential animats could create imitators out of the other animats, which would be detrimental to the tribe. It's great to have an animat that is really good at stacking bricks, but you don't want everybody to be stacking bricks. Did you see that kind of behavior in your simulations?

It might be interesting to test this by having individual "baby" animats rather than batch generations. Any chance you will release the code as open-source?

Great job! Seeing your defense gives me hope that my goals are within reach. I read your whole presentation. Thanks for posting it! We share a lot of interests!

I agree with Jim Clay about why Age authority might do better than merit authority.

Congrats Doc!

Thanks everyone! :)

Jim: Perhaps the slides alone weren't clear enough. You get the "co-commanders" (as you called the effect) in AH and MH social structures, the only difference is the ordering. The best explanation I can think of for why AH does better than MH is that, just as with humans, the agent that can do the job the best individually isn't always the one you want to put in charge. The efficiency with which an animat completes its own tasks is governed by its Actual Neural Network, but the quality of its commands depends on the Signal Neural Network, and they're reinforced very differently. The ANN is trained to maximize score, but the SNN is rewarded when it receives positive feedback from the animats it issued commands to. Animats that issue good commands won't necessarily have higher scores than others.

As for the neural networks being static in older animats, not really. I didn't use any sort of learning rate decay, and older animats can learn as quickly as youngers can.

Also, I think you misunderstand the "generation" system. Animats are created at the beginning of the scenario with staggered ages so that they die one at a time, and each time one dies it is replaces by a new "baby". There aren't batches of animats in a "generation".

Dallas: Oh yes, grad school is all about tenacity!

Congrats!

Hey, congrats!

In some systems, quality DOES rise to the top!

MC

Congratulations, Dr Michael!!!

Great job Dr. WIlliams. I don't have a clue what your dissertation is about. The comments above might as well be written in Sandskrit, but I'm sure you'll use what you've learned for good.

MC said: "quality DOES rise to the top"

All too infrequently.

congrats doc! Hope you took a well deserved rest this weekend :)