Universal Agents

Transcript

Hi! I'm Tim Tyler, and this is a video about Universal Agents.

I will also be covering the concepts of Universal Sensation, Universal Action, Universal Computation and Universal Memory.

A brief history of universality

Long ago, John von-Neumann introduced the ideas of universal computation and universal construction - in the context of cellular automata.

Universal computation turned out to be essentially equivalent to Turing completeness. It became a very popular and important concept. Computation universality eventually became widely shortened to universality, in cases where there are enough context cues to understand the abbreviation.

By contrast universal construction suffered from problems. It wasn't terribly obvious what it referred to in the case of irreversible automata, and it turned out not to map neatly onto the concept of being able to support self-reproducing creatures - as was illustrated by Chris Langford with his self-reproducing loop.

Cybernetics

Today, I will be continuing in the grand tradition initiated by von-Neumann.

However, I won't be working with cellular automata today. Instead we will be looking at cybernetic systems.

I will be introducing several new universal concepts.

In digital cybernetics, there's the concept of an agent. Agents repeatedly accept digital sensory inputs, and simultaneously produce digital motor outputs at discrete time steps.

Note that in this type of cybernetics, things like battery life, resource and storage limits, error detection and error correction are abstracted away and ignored.

Universal Sensation

Universal Sensation refers to the idea that agents can potentially experience and endure any finite sequence of sensory input symbols. They can enter states where aren't any specific sequences that cause them to explode, shut down, turn their sensors off, unplug themselves - or otherwise cease witnessing. The agents should be able to attend for unlimited periods of time.

Universal Action

Universal Action refers to the idea that agents can issue arbitrary finite sequences of motor outputs. There is no sequence of motor output actions that they cannot perform. The agents can't bash their own brains in. They can't turn themselves off, sabotage their motor outputs, or do anything that might interfere with their motor outputs being generated. Maybe they can chop their own actuators off - but there must still be some level at which their motor outputs can still be detected.

Universal Computation

Agents can also exhibit Universal Computation, which means what it usually does.

Universal Memory

There's also the idea of Universal Memory - which is the ability to remember any finite sequence of symbols.

Prerequisites / dependencies

Some of these concepts depend on other ones. The dependencies are as follows:

Universal Computation depends on Universal Memory;
~~Universal Action also depends on Universal Memory~~;

Universal Agents

So, those are the basic definitions. Why might one be interested in these concepts?

Universal Action affects whether an agent can perform destructive brain surgery on itself. Agents that can't do this tend to be immune from certain theoretical problems that plague intelligent machines - such as the wirehead problem.

Then there's the issue of whether agents can learn arbitrary action sequences in response to arbitrary sense data - and so become Universal Agents. The capabilities of Universal Computation, Universal Memory, Universal Sensation and Universal Action appear to be prerequisites for this type of Universal Learning.

I hope these concepts prove useful in thinking about the possibilities for synthesizing agents.

Enjoy,

Tim Tyler | Contact | http://matchingpennies.com/