Case-Based Reasoning (CBR) is the process of
intelligently solving new problems based on the previous similar problems. The
basic steps of CBR are retrieve, reuse, revise and retain.
First, CBR retrieves the most relevant case to the current problem at
hand. The retrieved case is reused and revised to incorporate minor variations
in the solutions. This adaptation step gives CBR a power to form more precise
and accurate solutions to the future problems. Finally the revised solution is
retained for future use.
The case was represented using ten parameters (the closest food, enemy and
obstacle, the density of food and enemies, the ratio between the agents health
and the opponents, and parameters determining the actions currently available
to the agent). A weighting scheme was used to emphasize the more important
features. We have identified 19 historic cases that are selected based on the
performance of agent in the training simulations. Even with the small number of
cases, the CBR-based agent shows drastic improvement over the random behavior
and can be refined by adding more specific cases to the library. The selection
of the case was done with a distance matrix based selection. The reliability of
the case based on the previous failures is used as a weight in order to
encourage the selection of variety of cases. The reliability of the case is
dynamically adapted depending on whether the case can or cannot be successfully
applied to the current situation.
|
|
After the action to be performed is selected by the CBR module, a set of
heuristics are used to adapt the action to the current environment. First, the
heuristic module checks whether the proposed action is doable. In case of
failure, the reliability measure of the case which proposed the action is
reduced. The reliability of the case is used as a weight during the calculation
of the distance matrix, so that case with the less reliability will have
greater distance in the subsequent cycles. The second heuristic calculates the
direction and the speed of the agent movement. For instance, eating a food item
requires the direction toward the food and appropriate setting of the speed
such that the agent stops at the food item.
|
