There are other genetic algorithm parameters which areimportant to RBGAF-SLAM. Mutation rate, mutation range,crossover rate… All these parameters have been properlychosen so that it works find in real implementations as shownin [11]. However, the affection of these parameters on othercomplex environment is not investigated. Therefore, it isnecessary to design the learning algorithm to learn theseparameters so that RBGAF-SLAM will be able to functionwell in various environments.Few major issues are addressed for RBGAF-SLAM abovein our experiments. These problems can be solved by increasethe number of Rao-Blackwellised particles. A large amount ofRao-Blackwellised particles significantly improves therobustness of the SLAM approach. However, due to thelimitation of computational resources on the robot, the numberof Rao-Blackwellised particles is strictly limited to tens tohundreds.However, the offline computational power is much morethan the processors on robots. Therefore, it would be nice ifoffline computation could ‘help’ the real-time SLAMalgorithms to perform better. In our approach, offline geneticlearning techniques are implemented to search for a proper setof RBGAF-SLAM parameters. Although the computationalcost is high, it does not have the ‘real-time’ restrictions.IV. GENETIC LEARNINGExpectation maximization (EM) strategy can be used to findout the optimum sets of parameters for the RBGAF-SLAMapproach. However, EM theory suffers from the localmaximum problem. The success implementation of the EMtheory is dependant on a good initial ‘guess’. A good initialdata set is needed. Due to the high complexity of the RBGAFSLAM problem, it is quite difficult to find a good initial dataset. Implementing EM theory for searching the properparameter set may lead to a success but quite risky. It mayprovide an un-acceptable sub-optimum solution and requireslots of testing/tuning to get the good initial data set. Treedecision and are another options. However, these methodsare more suitable to provide optimum solution in the discretespace instead of a continuous space.Genetic algorithm has been proven to be a good globaloptimization tool for a number of implementations. It does notrequire a good initial data set to find out the global optimum,although good initializations can often faster the search. Asdescribed in [13], GA is motivated by a number of factorsincluding:“Evolution is known to be a successful, robust method foradaptation within biological systems.Gas can search spaces of hypotheses containing complexinteracting parts, where the impact of each part on overallhypothesis fitness may be difficult to modelGenetic algorithms are easily parallelized and can takeadvantage of the decreasing cost of powerful computerhardware.”GA can deal with both continuous and discrete data. Theprocesses involved in genetic algorithms like re-production,selection and evaluation have been extensively investigated.Researchers have developed systematic GA implementationstrategies with lots of open source available for theimplementations. Due to these advantages, genetic learningapproach is selected for optimizing RBGAF-SLAM.To determine the GA techniques and correspondingparameters, we designed genetic learning approach onRBGAF-SLAM. The overall design of the genetic learning isshown in Fig 2.