Reinforcement learning. 2.4.1 Basic concepts Reinforcement learning (RL) is learning what actions should be taken to maximize the a numerical reward [110]. In the most interesting and challenging cases, actions taken in the current state may affect not only the instantaneous reward but also the next state, and thus all subsequent rewards. Two characteristics, namely Trial-and- error search and delayed reward, are the most significant distinguishing features of RL. The problem of RL is usually formalized under an incompletely-known Markov decision process, and the basic idea of RL is that the learner interacts with the environment according to the behavior policy to update the target policy. In RL, we often meet two terminologies, namely, on-policy and off-policy. For on-policy, we have the same behavior policy and target policy. For off-policy, the behavior policy and target policy are different. To compare RL with the most popular categories in current machine learning research field, i.e., supervised learning and unsupervised learning, we give a general review of them, as follows • Supervised learning learns from a training set of labeled examples to model relationships and dependencies between the target prediction output and the input features [92]. In the past decades, a wide range of supervised learning algorithms have been developed, such as linear regression, logistic regression, support-vector machines, K-nearest neighbour algorithm and Naive Bayes. However, it is not adequate for learning from interaction because the agent would be expected to learn from its own experience in uncharted territory. • Unsupervised learning is typically about finding the pattern/structure that are hidden in the collections of unlabeled data [53]. Many superior unsupervised learning algorithms, e.g., K-means and principal component analysis, have been developed and applied to our real life. K-means clustering methods automatically groups its training examples into categories with similar features [127]. Principal component analysis algorithm is to compress the training data set via identifying useful features and discarding the rest [124]. Unlike unsupervised learning, RL is trying to maximize the numerical reward instead of finding hidden pattern/structure.
Reinforcement learning. An Introduction. The MIT Press, 1 edition, 11 2017. complete draft. [106] United nations of roma victrix (unrv), roman taxes. xxxxx://xxx.xxxx.xxx/ economy/roman-taxes.php. [107] L. xxx xxx Xxxxxx, X. Postma, and H. xxx xxx Xxxxx. Dimensionality reduction: A comparative review. Technical report, Maastricht University, 2009. [108] F. xxx Xxxxxx, X.X.; Xxxxxxxx. Bpi challenge 2018. xxxxx://xxx.xxx/10. 4121/uuid:3301445f-95e8-4ff0-98a4-901f1f204972, 2018. [109] S. F. Wamba and X. Xxxxx. Big data analytics for supply chain management: A literature review and research agenda. In Workshop on Enterprise and Organ- izational Modeling and Simulation, pages 61–72. Springer, 2015. [110] X. Xxxxxx. Investeringsagenda belastingdienst. xxxxx://xxx.xxxxxxxxxxx. nl/kamerstukken/detail?id=2015Z09033&did=2015D18368, 2015. [111] Wikipedia, English edition, history of statistics. xxxxx://xx.xxxxxxxxx.xxx/ wiki/History_of_statistics#Etymology. 140 Bibliography [112] Wikipedia, English edition, trial of the pyx. xxxxx://xx.xxxxxxxxx.xxx/ wiki/Trial_of_the_Pyx. [113] R.-X. Xx, X.-S. Xx, X.-x. Xxx, X.-X. Xxxxx, and X. X. Xxx. Using data mining technique to enhance tax evasion detection performance. Expert Systems with Applications, 39(10):8769–8777, 2012. [114] X. Xx, X. Xxxxx, X. Xxxxxxx, X. X. Xxxxxx, and N. V. Chawla. Detecting anomalies in sequential data with higher-order networks. arXiv preprint arXiv:1712.09658, 2017. [115] X. Xxxxx, X. Xxxxx, and X. XxXxxx. The Practice of Statistics. X.X. Xxxxxxx, New York, 1 edition, 1999. [116] X. X. Xxxx. Spade: An efficient algorithm for mining frequent sequences. Ma- chine learning, 42(1-2):31–60, 2001.
