Agent programs

All agent programs that will be presented on this web site have the same structure: they take the current perception results from the sensors as input data and return the selected action to the executive mechanisms6. It is necessary to point out the difference between the agent's program, which accepts the results of the current perception as input, and the agent function, which accepts the entire history of acts of perception at the entrance. The agent's program receives as input only the results of the current perception, since it can no longer learn anything from its environment; if the actions of the agent depend on the entire sequence of acts of perception, then the agent must himself remember the results of these acts of perception.

A simple pseudocode language will be used to describe agent programs. For example, the following listing shows a fairly simple agent program that registers a sequence of acts of perception, and then uses the resulting sequence to access by index the action table and determine what needs to be done.

The table clearly displays the agent function implemented by this agent program. To create a rational agent in this way, designers must create a table that contains the appropriate action for any possible
sequence of acts of perception.

function Table-Driven-Agent ( percept ) returns action action
static: percepts , sequence of acts of perception,
originally empty
table , action table indexed by
sequences of acts of perception and
fully specified from the very beginning
add percept percept results to end
percepts sequences
action <- Lookup ( percepts , table )
return action

Consider an automated taxi: visual input from one camera comes at about 27 megabytes per second (30 frames per second, 640x480 pixels with 24 bits of color information). According to these data, a search table calculated for 1 hour of driving must contain a number of entries exceeding ^{10,250,000,000} . And even a search table for chess (a tiny, well-studied fragment of the real world) would have at least ^10,150 entries. The staggering size of these tables (while the number of atoms in the observable universe does not exceed 10 ⁸⁰ ) means that, firstly, no physical agent in our universe has space to store such a table, and second, the designer cannot find enough time to create this table, thirdly, no agent will ever be able to learn what is contained in all the correct entries of this table, based on his own experience, and, fourthly, even if the environment is simple enough to be created tabl zu acceptable size, it is still the designer no guidance information about how to complete the recording of such a table.

Despite all this, the Table-Driven-Agent program does exactly what is required of it: it implements the desired agent function. The main difficulty facing artificial intelligence as a scientific field is to learn how to create programs that, as far as possible, produce rational behavior using a small amount of code, rather than a large number of table entries. There are many examples showing that such a task can be accomplished successfully in other areas; for example, the huge tables of square roots used by engineers and schoolchildren until the 1970s are now replaced by a five-line program in electronic calculators that uses the Newton method. The question is, can artificial intelligence do for intellectual behavior in general what Newton did to simplify the calculation of square roots? Many scientists believe that the answer to this question is positive.

The rest of this section covers four main types of agent programs that embody the principles underlying almost all intelligent systems:

• simple reflex agents;
• reflex agents based on the model;
• agents acting on a goal basis;
• agents acting on the basis of utility.