How can we develop and test an Artificial Intelligence system? - Artificial Intelligence

When building any system, an engineer must face the question of how best to evaluate how effective it is. If engineers ignore the evaluation phase, then they are in danger of make the same mistakes that early AI researchers did when they often omitted a full evaluation of their system. (This is often a mistake made with current AI systems as well). Evaluation is more than just testing the system to see if it is working correctly. Testing is a necessary step in the software engineering process that should not be omitted. However, evaluation goes further, and analyzes the results produced by the system, and compares the results with those produced by other systems. Without this analysis, then it is impossible to ascertain whether true progress has been made. That is, whether there have been improvements in whatever criteria are being used to evaluate the systems, or whether the design objectives have been met (such as the design objectives proposed in Chapter ).

How, then, can we develop and evaluate an AI system? One way is to create a virtual environment with as much detail and complexity that the real world has, not unlike the environments that the Catching Features software is able to create, and have the A.I. physically grounded within it. There are two advantages to this:

  1. The environment is computer generated. That is, the computer already knows everything about it, since it has already been generated, and there are data structures that represent every aspect of it. In a real sense, we can neatly sidestep the issue of knowledge representation .

  2. If the virtual environment can be visualized, then both the human and A.I. can interact together within the same environment.This could be useful if we wish to take the role of teachers to teach the AI agents to aid them in learning about the world and how best to interact with it. The role that the teacher takes could be non-competitive – more as a benevolent tutor who oversees the learning process – or it could be competitive as in the orienteering simulation where the goal is to get the computer agent to reproduce the behaviour of the human players and attain the same level of performance or perhaps exceed it. If the latter is possible, then the computer agent can take over the role of teaching or helping the human players perform better themselves (this may simply happen by the human trying to compete at the same level as the computer agents).

All rights reserved © 2020 Wisdom IT Services India Pvt. Ltd Protection Status

Artificial Intelligence Topics