Principles of intelligence

- Research Unit 2 -

A central objective of Science of Intelligence is to identify the principles of intelligence. These principles must “explain all instances of intelligent behavior, i.e. in our case: all behaviors synthesized and analyzed in Research Unit 1.

What are principles of intelligence in SCIoI?

Principles represent regularities that remain valid across all observed instances of intelligent behavior. These regularities describe how behavior must be tailored to the characteristics of an ecological niche to be adaptable, general, cost-effective, and goal-directed in the real world (see our preliminary definition of intelligence). For example, in some ecological niches, behavior is best generated by an individual agent, in other niches collective intelligence is a better approach, and in a third niche the two must be combined. A principle describes this relationship between behavior generation and niche and can thus be used to synthesize an intelligent agent or to analyze and explain its behavior.

Principles could hence be general rules that make it possible to adapt the parameters, architecture, or algorithms of agents to their respective ecological niche. Principles could also have parameters themselves. This will be likely for behaviors that are subject to trade-offs (exploration vs. exploitation, nature vs. nurture, modularity vs. integration). These trade-offs can take the form of hyperparameters for principles that optimize an agent’s parameters (e.g. discount factors for reinforcement learning). On a higher level, a principle could also be a rule how to choose the optimal trade-off, given the nature of the agent and its ecological niche.

How can we identify candidates for principles of intelligence?

In SCIoI, research on the example behaviors in the Research Unit 1 will serve both as an incubator for principles of intelligence and as a testbed for candidate principles. While each of the behaviors synthesized in Research Unit 1 faces its individual challenges, some are common to several if not all of them. For example, each behavior requires agents to select from the environment information that is informative for its future actions. Each behavior requires agents to develop a model of the environmental components it aims to manipulate, be they objects, other agents or the collective behavior of a swarm. The individual behaviors can hence provide candidates for general principles in a “bottom-up” manner in the form of solutions to problems that re-appear in other intelligent behaviors. Candidates for principles will thus serve as a cohesive force between the different example behaviors.

Candidates for principles could also be identified in a deductive top-down manner by developing solutions to problems that are intrinsic to a broad spectrum of intelligent behaviors. Here, the research questions will serve as a starting point. For two of these research questions, we outline examples of such top-down approaches below (curse of dimensionality and modularity vs integration). Again, candidate principles should be beneficial to all example behaviors.

During the initial research phase, top-down candidates for principles will be developed and then tested in the different example behaviors. They will increasingly be complemented by “bottom-up” suggestions as research on the example behaviors advances.