3. Omniscience, training and autonomy
It is necessary to pay attention by distinguishing rationality
from omniscience. An omniscient agent knows the result of its action and acts
as consequence. But actually omniscience is not possible.
Rationality is different from the perfection. Rationality
maximizes the awaited output, whereas the perfection maximizes the real result.
Thus, the definition suggested for rationality does not
require omniscience, since the rational choice depends only on the continuation
of perceptions received until a given date.
A rational agent is also that which learns the possible
maximum of what it perceives. The initial configuration of an agent can reflect
a preliminary knowledge of the environment, but as the agent acquires the
experiment, its knowledge can change and increase. There are exceptional cases
where one knows the environment completely a priori. In these cases, the agent
does not need to perceive nor to learn; it acts simply in a correct way. But
these agents are very fragile.
An agent misses autonomy when it is based more on its initial
knowledge which its originator provides him than on his own perceptions. A
rational agent must be autonomous, must know to learn how to determine, how to
compensate for an incomplete or initially partial knowledge.
In practice, it is little of time that a complete autonomy is
required as of the beginning: when the agent has little or no experiment, it
will have to act in a random way unless the originator does not provide him
with any help.
| 
