This paper presents a computational model
capable of improving the action policies for a well-defined
domain. Each action policy is represented as a driving plan P,
which is composed of a number of actions {a1, ... an}. These
actions can be used to move a train in a stretch of railroad Sti.
The plans are elaborated using a CBR approach and reusing
previous solutions and learning from plans. The CBR cycle is
divided in three agents that act collaboratively to build or run P,
namely: Planner, Executor, and Memory. The Planner agent
generates P. Executor agent is responsible for revise and apply
the actions of P. During the execution, P may undergo several
adjustments depending on multiple circumstances, such as
environmental conditions. The modified plan P’ returns to its
origin end to integrate the local case base, managed by the
Memory agent. This approach was evaluated on the following
criteria: (i) fuel consumption, (ii) accuracy of case retrieval, and
(iii) efficiency of adaptation task and application of adapted
cases in real-world scenarios. The inclusion of new experiences
reduced efforts Planner and Executor in their tasks, and a
reduction in fuel consumption. In addition, the model shows that
the increase in diversity in the case base increases the reuse of
experiences in objectives-scenarios.