GoID
Source and stuff for the Game of Intelligent Design
Overview
GoID (or the Game of Intelligent Design) is an exploration of artificial general intelligence (AGI) concepts. Originally it even was a game, but its design didn't foster a community the way that it needed to. It used to be a set of tasks, and you would write an agent (in Javascript) to perform each. After the agent finished the task it would be scored on criteria intrinsic to the specific task. (The code for the original game will be added to this repo for historical interest.)
This time, GoID is a single task that will get more and more complex over time. The agents that people write will of course have to become more complex as well. You might consider this a form of agent evolution. This is where the software gets its name: players are the intelligent designers of the agents that, with hope and a diaper load of luck, will become the basis of AGI. At the very least we might expect to shed some light on how nervous systems work, how data is represented in brains, or any of the other countless AGI puzzles that have yet to be solved.
Core concepts
The study of artificial intelligence started in the 1950s. Over 60 years later we still don't have any of the technologies that people back then didn't seem to think were going to be a big deal to create. Sure, computers can play chess and Jeopardy!, but, as one researcher pointed out, they don't know to come in from the rain. People will say that there is artificial intelligence at work everywhere, from driving cars to writing hotel reviews to finding me movies to watch. But these are all examples of "narrow intelligence", or fancy software that is very good at doing one very specific task, but is incapable of anything else. Humans, the only example of general intelligence in existence that we know of, are capable of much, much more. Except, it seems, to be able to build another general intelligence.
This, however, is not for lack of trying. Over those 60 years there have been all manner of attempts to create intelligence, from symbolic processing to artificial neural networks (ANNs) to fuzzy logic to deep learning to predictive analytics to creating detailed simulations of neural physiology to the hundreds of other approaches, most of which i haven't heard of or understand. And it's not like the people who tried these things were not qualified or anything; we're talking about some very big-brained people. So what's been going wrong? Well, i don't know the answer to that question, but i suggest it doesn't matter. (Ok, in some cases i know what is going wrong, but whatevs, it still doesn't matter.) What we need is an entirely new approach. If what we want is something like a human intelligence, we need to create it in a way that is similar to how it was created in the first place: evolution.
Ok, yes, this has been tried too - many times - but small changes in the approach can have important consequences, and i believe that what i am proposing is truly different. Also, if it's clear that there is a problem with the approach, it can be adjusted.
The core concept is that of agent situation: the agent has to live in a simulation that is as much like our world as is practical. To start, i will propose a simulation that will be disappointingly simple for some, but we have to start somewhere, and the intention is for it to become more complex over time.
It is important to remember that there is only one simulation environment, but there can be any number of agent implementations that can run in it. (In the future, multiple agents will be able to run at the same time and compete.) We're not after "the right answer" here, but rather creating a place where many ideas can be explored at the same time.
Other important concepts help define the approach. One is temporal resolution: the agent lives in real time, and receives sensory input at the millisecond level (somewhere around every 10 or 20 milliseconds, which nicely matches the frame rate of simulation viewers, and is in fact similar to the firing rate of neurons).
A more advanced concept is biologically-inspired movement. You don't "turn left", "turn right", "move one block forward"... You apply a force to a joint over time like a muscle would. I call this more advanced because force-based movement (as opposed to inverse kinematics) is maddeningly difficult to get working well. (See https://www.youtube.com/watch?v=K5DcK-ImFxc for an attempt to create an artificial inchworm.) This will not be in the original simulation, but will be eventually because it's killing me to understand it.
Details
Specifically, there is an Environment, in which an Agent lives. (Later, multiple agents.) The agent's body is part of the environment; the part that is really the "agent" is what controls it, which we can call the Brain. At its simplest, the brain receives, say every 10ms, a set of values from the environment, which we'll call its Sensory Input. It can do with this input whatever it likes: process it, store it, ignore it... While it does this, it sets values in its Actuators, or outputs, which typically will apply a force to a joint that causes the agent's body to move at the end of the cycle. The agent body is subject to the environment just as real-world living things would be. If the agent's brain does not take actions to do things like eat or hide, it will eventually end up dying, which is what we're trying to avoid.
Over time we make the environment more and more complex, which forces the brains to do the same. What we want is to gradually build up a biologically plausible primitive brain structure that, hopefully, will answer some basic questions about intelligent brains, such as how memories are likely stored, how brains represent and process data, how motor planning works, maybe what the basic function of a cortical column is. These are lofty goals, but even if the project ends up only giving good ideas to smart people, it has a good chance of getting the world that much closer to AGI than it was before. And that, my friends, is the point.
So, who's ready to play GoID?
(Get it? Intelligent Design... playing GoID... It works right?)