Information Technology Reference
In-Depth Information
longest K-line that leads to a human-win. In other words, the computer player tries to
minimize its chances for defeat, and in doing so it looks for a move that - from past
experience, lead to either a computer win, or (if that is not available) to a tie, or (if that
is also not available) to a move that will postpone the defeat of the computer player the
longest. Assume K is the K-node that is selected for the computer player to move,
based on the above conditions/criteria. Then, the computer player moves by making the
move dictated by the found K-node K . Note, at this point an intersection is formed in
the AKL. Namely, the current K-line L2 intersects with the K-line containing the found
K-node K . The game resumes, with the human player making the next move. After
that, the computer player moves and its move is, again, based on the criteria described
above - i.e., by finding (if available) a K-node that leads to an educated guess of post-
poning its defeat for the longest. The process continues, as described above, and new K-
nodes are formed, as applicable, K-line intersections are also formed (as described
above), and the AKL is taking shape, for as long as it is desired.
Testing. In testing our system, we judge the computer's learning progress by monitor-
ing the amount of effort that is progressively required by the human to win. As the
AKL becomes denser, it is expected that the computer player will choose K-nodes
that belong to winning K-lines and, as a result, the moves that correspond to those K-
nodes will make it harder for the human player to win. Of course, for the simple game
of Tic-Tac-To, the human can always win, or, in the worst case, force a tie. However,
for a more complicated game such as chess (or go), it is not at all clear that the human
can keep winning after several games have been played. We expect that with enough
training of an AKL, the computer can become a better expert in any of these complex
games. We intent to test our method and report our findings, in a sequel paper.
Why is this way of machine learning interesting. Our proposed machine learning
method (via the formation and utilization of an AKL) is novel and interesting because
it does not require that any knowledge is available up front . The AKL can be formed
progressively, as more games are played (i.e., more knowledge and experience is
acquired), and there is also no limit of how much knowledge can be accumulated. As
more knowledge and experience is acquired and captured into the AKL, the computer
learner's performance is expected to improve. This, we argue, resembles the way that
actual human beings learn. That is, as more experiences are acquired with the passage
of time, humans, in general, perform better. Another interesting point in the nature of
the learning facilitated by the AKL is that the quality of learning is proportional to
the quality of the “teacher” . It is expected that an AKL whose K-node are formed by
moves that are made by a not-so-intelligent (or, not-so-careful) human being, will be
inferior to an AKL whose K-nodes are formed by moves that are made by a highly
intelligent (or, skillful, or very careful) human being. As such, a machine that bases
its moves on an inferior AKL is expected to underperform a machine that bases it
moves on the superior AKL. Note, this analogy also resembles the real-life scenario in
which, in general (all other things being equal), students of better teachers tent to
become more skillful (i.e., learn more) than students of not-so-good teachers.
Search WWH ::

Custom Search