A　method　of　Upper　Limb　Activities　Recognition　(UPLA)　based　on　Neural　Networks　is　presented.　The　accuracy　of　activity　recognition　will　be　influenced　by　the　size　of　sliding　window,　the　overlapping　of　adjacent　sequences　and　the　number　of　neurons　for　neural　networks.　Whereas,　there　is　less　work　in　hyper　parameters　optimization　of　neural　networks　automatically.　It　is　very　time-consuming　to　optimize　hyper　parameters　by　experts　through　an　experience　and　error　approach.　In　the　paper,　Genetic　algorithm　is　used　to　find　the　best　hyper　parameters　automatically:　the　size　of　sliding　window,　the　overlapping　of　adjacent　sequences　and　the　number　of　neurons　for　neural　networks.　The　basic　genetic　algorithm　has　a　slow　convergence　problem　and　it　is　very　easy　to　fall　into　a　local　optimum.　To　solve　the　problem,　the　population　selection　mechanism　is　improved.　A　comparison　is　made　for　the　improving　method　with　seven　traditional　classification　algorithms　and　convolutional　neural　network,　an　accuracy　of　97.9%　is　reached　by　using　the　new　method.　Finally,　an　App　is　developed　that　can　collect　and　recognize　upper　limb　activity　in　real　time.