The Learning Agent is another method of computer learning used in artificial intelligence. It uses an entity, called an agent which can interact with the environment and learn from the outcome. In our Blackjack situation, six players act as agents with a shared memory for their actions and responses. The tally statistics serve this purpose.
For each hand and dealer card combination, the actions are each tried in turn for the specified number of times (set by the user). First standing is tried, after which doubling down is tried, followed by drawing. If one of these actions is found to be profitable then the remaining action(s) are not evaluated. If none make money then after the performance data for all three actions has been compiled, the action that mitigates the losses is taken from then on.
In a manner similar to the other actions, insurance and splitting pairs are “learned” by first declining for the specified number of tries. If this alternative is not a money maker then taking insurance or splitting for the same number of times is tried. Afterwards the alternative that makes the most money is selected from then on. But, insurance and splitting pairs are dependant on the hands being played correctly in order to obtain a valid result. For this reason the “Learning Agent” dialog box has a check box for you to tell the program to start learning insurance and splitting. This way you can wait until the drawing, standing, and doubling is learned reasonably well first.
Also in the dialog box are check boxes to tell the program to relearn various actions. To relearn an action, the tally statistics for that action are reset to zero and the action is tried a set number of tries again. This feature can be used, for instance, to see if drawing were tested again would the tally statistics show the same conclusion. If the number of tries is small there is a greater possibility that the results may not be representative of the true statistical outcome. If the number of tries is large, the results are more statistically valid, but the simulation will take longer to run. Individual hand and dealer combinations can also be chosen to be relearned in the tally window. After clicking the Erase check box, select which squares are to be set to zero and thus relearned.
The Learning Agent mode will evaluate alternatives till it finds those that make money or minimize losses. Use the Tally Window to examine the statistical outcomes obtained. See how well these compare to the Basic strategy. In general, a table of six players play till the desired number of hands is reached. At this point the statistics for the best player is shown on the screen which allows the ending balance to be monitored to see how it improves over time. The balances are then reset to $500 and the deck is shuffled as the next table starts. The simulation will continue till stopped by the user.
