The following is comprised of a collection of on-going research into the options of AI approaches and specific architectures that fit my previous criteria as closely as possible. Theoretically, any AI architecture or method could be considered here, but that it would be a huge waste of time to elaborate on them all. Instead, I’ve selected to discuss candidates that I know fit at least most of my criteria. And I know that I will need more (detailed) research here not only in the architectures themselves, but how people author for them.
First, a review of terminology that has caused many lines to be drawn in the sand of the AI landscape. Here are some AI keywords (from a surfing of wikipedia) that will narrow down the field of all “artificial intelligence” to those close to my specifications:
Situated Artificial Intelligence
While situated often implies robots in the case of this field of AI, it has been accepted that virtual agents embedded in a dynamic virtual world that they can sense and manipulate also count as situated. I suppose that situated may more closely represent what I mean by embodied. Situated AI tends to be represented bottom-up: composing tiny bits of behavior into less-tiny-bits of behavior in hierarchical steps until we reach a composition that will appear to look like “opening a door” to a viewer.
Traditional or Symbolic AI
I include this definition in contrast to the last. It is top-down: decomposing the idea of “opening a door” into many individual sub-steps until a list of actions are found. Static Behavior Trees (BTs), Finite State Machines (FSMs) are included in this class of behaviors. It is impossible to foresee all situations an embodied agent will encounter, and traditional AI falls short when the agent encounters an unknown/missing/inappropriate symbol or behavior.
Psychology’s affect deals with emotions and feelings. AI in this field process audio and visual signals relating to human reading and demonstration of a variety of emotions (which, in themselves, are difficult to classify). My concern with emotions is enabling the author to create agents that dramatically express (and possibly track) emotional state in a human-readable fashion. It would be a first-class research problem to try and make an AI try to read the player’s emotions (which is why I am dodging the input/output problems).
Artificial General Intelligence
While an author could make an animal or non-human agent enact their behaviors, we are aiming to author human-level intelligence in AI. However, we want human-level intelligence in the service of human-like performance. AGI is hypothetical and controversial; are we creating intelligent agents (AI-hard) or agents that act intelligent? I am perfectly satisfied with simply acting intelligent.
Automated Planning and Scheduling
Often shortened to just “planning,” this branch of AI is concerned with composing a sequence of decisions, often in service of accomplishing a goal. Planning can go forward and backwards and be driven by any manner of decision-making theories such as reinforcement learning and statistical models. Our agent will need to have some method of planning in its architecture for sequencing and deconstructing behaviors (so that they are simpler to author).
Brains and computers have shared a metaphor since computers were invented. We can use our own mind’s inputs, outputs, and computations as inspiration for computational models, and once we make them accurate enough, we’ll have solved that hard AI problem. Theoretically. Until then, we can only work with the closest approximations we have. If a cognitive science system can produce decent behavioral performances through its approximations, it deserves to be considered as another model of decision-making.
Next post will be about actual architectures. Promise!