Representation

Learning has proved possible -so far- only where a representation of the problem and all its candidate solutions can be found that is simple enough to constitute a tractable search space for computer algorithms.

The human mind can make on-the-fly analysis of huge amounts of information. The raw state of the game at any given instant has some 232 bits of information -too much for rapid analysis. Instead, artificial Tron players have been provided with 8 radial distance sensors, a tractable source of information.

The behavior of artificial Tron players is controlled by a `genotype' consisting of 32 integer numbers. Each value represents the relative weight that a sensor has in each of the four possible directions that can be chosen.

At every game step the player will go to the direction with a highest score in terms of the current sensor readings and its individual weights.

Sensory model: Eight sensors are the
information source available to artificial
Tron players.

Human programming: Given this representation,
it is not hard for a human to create fairly
good players. In this example, the program
Avoider (blue), whose strategy is staying
away from obstacles, is playing with a human
opponent (black).


NEXT / PREVIOUS / INDEX / HOME
PLEASE SEND YOUR COMMENTS! pablo@cs.brandeis.edu