We investigate the use of Stochastic Local Search (SLS) technique to explore environments where agents? knowledge and the time to explore such environments are limited. We extend a work that uses evolutionary algorithms to evolve teams in simulated environments. Our work proposes a formalization of the concept of state and neighborhood for SLS and provides evaluation of agents? teams using number of interesting cells. Further, we modify the environments to include goals that are randomly distributed among interesting cells. Agents in this case are then required to search for goals. Experiments using teams of different sizes show the effectiveness of our technique. Teams were able to complete exploration of more than 70% of the environments, while in the best cases, they were able to complete explorations of more than 80% of the environments within limited time steps. These results compare with those of the previous work. It is interesting to note that all teams of agents were able to find on average all the goals in the three environments when the size of the grid is 12. This is a 100% achievement by the agents? teams. However, performance can be seen to degrade as the environments? sizes become larger.
Part of the book: Artificial Intelligence