The invention discloses a personalized exercise recommendation method fusing reinforcement and comparative learning, and the method employs a self-supervised
reinforcement learning method, enables an exercise recommendation process to be formalized into a Markov
decision process, carries out the interaction between a student and an
intelligent agent, and obtains a feature representation. Meanwhile, a comparative learning thought is integrated, similar learning histories are drawn together, and different learning histories are separated, so that effective characterization is learned; splicing the hidden states to obtain spliced representations # imgabs0 # and # imgabs1 #, and further obtaining a whole exercise and
interaction history representation; # imgabs2 # and # imgabs3 # are used as input, and a final Q value is obtained through calculation and splicing of a Q value function; then, through a self-supervised
reinforcement learning method, the accumulated reward is maximized; and in combination with self-supervision, reinforcement and learning loss comparison, model training is performed by using
gradient descent, a final exercise recommendation result is obtained by using the model, and more suitable exercises are recommended to students. The invention also discloses a
system for realizing the method.