Research On Multi-source Transfer Learning Algorithm For Lifelong Learning Agent

For problems such as path planning, text classification, face recognition, colorenhancement and optimal decisions, lifelong learning Agent will inevitably encountermassive data processing, deficient target training samples, high costs and multiplyrepeated tasks in intelligent optimization, machine learning, pattern recognition andimage processing. Regarding the above features of lifelong learning Agent, the papertries to apply the following research methods to achieve multi-source transfer learningin different fields:1. In machine learning, lack of training samples in classification prediction can leadto accuracy drop, hence, MSTDT method is proposed. At first, it will determine thesimilarity among decision trees by automatically selecting component probability orpath probability; secondly, it can choose whether to use multi-source integratedtransfer based on multi-source conditions. At the same time an extremely low targettraining samples or only one available sample shall be taken into account to analyzeface identification. At last it puts forward multi-source transfer algorithm based onLPP characteristic mapping and verifies the identification by using typical faceidentification database like FERET, ORL and Yale.2. An ELM multi-source transfer Q learning algorithm is brought forward whenreinforcement learning faces large scale or continuous complex curse ofdimensionality problems. ELM ensures the approximation of Q value function, whilethe multi-source transfer mechanism can reduce decision difficulty of target problems.In fact, the nature of transfer is the similarity measurement between task space andsample space, and by using prior probability one can ensure transfer task and sampleplay a positive role in the targeted task and prevent negative transfer from occurring.3. In color processing, due to color distortion caused by color sequence ambiguityand uncertainty, the paper proposes a multi-source color transfer algorithm based onactive profile exploration. It uses active evolution methods to generate virtual contourand applies energy function evaluation mechanism to force it is gradually approachingactual contour. Meanwhile consideration is taken for the expression, split andconversion of source and target images in different color spaces such as RGB, Grayand LMS to achieve its multi-source color transfer in l space. The comparisonand gray color channel selection tests of single and multi-source transfer prove thereasonability and effectiveness of the algorithm. 4. Intelligent optimized algorithm has different computational complexity withexponential growth and dependence on its multivariable coupling parameter settings,so the paper proposes Multi-Source Transfer Ant-Q and multi-source parametertransfer algorithm based on graph construction. The former analyzes the similarityratio between source and target tasks and determine each transfer samples by thisweight; the later constructs the model transfer graph of the source task includingknowledge (ACO operating parameters) to approximate the manifold space ofmultivariate parameters.