Unsuccessful agents

Since agents are a structure, it is possible to put an agent in an environment in which it is not possible for that agent to perform optimization. This is an edge case in an otherwise robust generalization that agents are a type of optimizer.

It is common in the related literature to focus entirely on behavioral definitions of agents, to the point where an “unsuccessful agent” would feel totally contradictory to some researchers. We believe that if you take a real implementation of, say, a utility maximization algorithm, and place it in an environment like those below, then it should still be classified as an agent and detected by formal definitions.

There are many types of environment that could prevent an agent from achieving its values. The agent could;

• be totally separated from its environment, with no input or output.
• have lots of observation and action capacity, but happen to be blind in the ways that are relevant to its values.
• have an impoverished model class, and be placed in an environment which cannot be effectively approximated within its model class.
• have been placed in an adversarial environment, one which deduces what the agent is trying to achieve, and makes changes to the environment to do the opposite.