Foundations of intelligence
What are the foundation of intelligence? What can make an artificial entity capable of intelligent behavior in its own environment? Here I list the major building blocks that are necessary to learn to solve multiple tasks while acquiring knowledge and material gains. This list exists to understand where the current status of research and developement is and where we need to focus.
—
An intelligent agent lives in an environment that can be observed, and the agent actions have an effect on the environment (the loop)
The agent has a purpose to conduct itself in the environment
A purpose is a collection of goals aimed at obtaining something from the environment
These goals can be: stay alive, learn to predict the environment and effect of actions, learn to predict other agents, all while gathering something material (resources) or immaterial (knowledge, information, feelings or collection of sensations)
The environment provides both positive and negative compensation: reward or punishment
The agent has a module to perceive compensation and evaluate it
This module is a learning module that modifies the agent behavior
Let us call this module a goal function module (GFM)
Learning means solving more and more complex tasks, or just a long series of tasks, many repeating very often
GFM is not just one function, it is a complex machinery — it is a predictive system that predicts the state of the environment after performing one of multiple action from current state
GFM is likely a collections of neural networks or similar learning systems
GFM is not just sequence imitation, or reinforcement learning, it is both and more
More because the agent needs to have a sense of accomplishment to learn new tasks and methods that were not in the example set
A curriculum is generally desired to learn more and more complex tasks
GFM needs data to learn, and to be able to learn sequences of observations, actions, rewards
GFM can learn one task at a time, but rarely has a chance: the default is learning multiple tasks in parallel
The environment rarely offers the ability to focus on one task and one set of data, most times the episodes and tasks are randomized
GFM thus needs to be able to do continuous and life-long learning
The most interesting aspect of GFM is that it learns automatically new GSM modules, it grows during learning
GFM thus needs to have enough learning capacity to begin with, as in most adult animals, or have the ability to grow more modules
GFM needs to be able to modify the agent goals in response to more learning
Interesting references:
One-Shot Imitation Learning
Imitation learning has been commonly applied to solve different tasks in isolation. This usually requires either…
arxiv.org
Learning One-Shot Imitation from Humans without Humans
Humans can naturally learn to execute a new task by seeing it performed by other individuals once, and then reproduce…
arxiv.org
Matching Networks for One Shot Learning
Learning from a few examples remains a key challenge in machine learning. Despite recent advances in important domains…
arxiv.org
robotology-playground/pybullet-robot-envs
pybullet-robot-envs is a Python package that collects robotic environments based on the PyBullet simulator, suitable to…
github.com