Research On Prediction With Confidence And Security Games For Decision-making Under Incomplete Information

There are many decision-making scenarios in the real world,such as assisted driving,medical diagnosis,financial forecasting and uav cluster confrontation.With the progress of information technology such as artificial intelligence and big data and the application of mathematical theories such as statistics and game theory,a large number of models and algorithms emerge and play an important supporting or leading role in various decision-making scenarios.For decision-making problems in social,science,economics,etc,the quality and reliability of decisions are crucial.Considering all kinds of uncertain and unknowable factors in the actual decision-making scenarios,it has important theoretical research value and practical application significance to study the decision-making problems under incomplete information.When the neural networks model is used to aid decision-making,the model’s prediction may not be trusted due to the influence of various incomplete information factors,so it is necessary to obtain the model’s confidence to prediction as a part of decision support information.In addition,for the security games problem that belongs to autonomous decision-making,existing researches have not fully considered all kinds of incomplete information factors,and not paid attention to the cost of resource scheduling in actual scenarios,which results in that the effectiveness of defense strategies in practical application cannot be guaranteed.In view of the above problems,the specific work of this dissertation is as follows:(1)For the prediction problem with confidence under incomplete information,this dissertation focuses on the regression prediction with confidence information(confidence degree and confidence interval).In this dissertation,"multiplicative Gaussian noise" technique in stochastic regularization techniques is taken as an example,which is applied to feed-forward neural networks,and the optimization objective based on stochastic regularization techniques is derived.Specific conditions are then given to make this optimization objective equal to it of Bayesian neural networks that uses variational inference.Thus,on the premise of no increase in the number of parameters,the feed-forward neural networks using stochastic regularization techniques can be equivalent to the Bayesian neural networks,and the output of the model is the probability distribution of the predicted value,in which the confidence degree and confidence interval can reflect the model’s confidence in the predicted results.(2)For the problem of security games with incomplete information,this dissertation focuses on repeated security games with multiple rounds.This dissertation proposes new performance criteria "reallocation times" and "reallocation quantity" in addition to the traditional "regret".At the same time,this dissertation proposes a new algorithm:Random-Walk Perturbations with Uniform Exploration Algorithm.The algorithm maintains a cumulative reward estimator for every target to be protected,and gives consideration to exploration and exploitation at every round of game,that is,the algorithm performs the exploration strategy with a certain probability,or performs the random-walk perturbation strategy.Then,the geometric resampling algorithm is used to calculate each target’s rewards estimator in the current round,for updating cumulative reward estimator of every target.This algorithm can be used to generate the strategy with both low regret and low resource reallocation for defender.In this dissertation,it is proved by experiments that feed-forward neural networks using stochastic regularization techniques can give the probability distribution of prediction,from which the model’s confidence to prediction can be obtained,which can be used as a part of decision support information and is helpful to improve the quality and reliability of decision-making.In addition,through detailed theoretical analysis and derivation,it is proved theoretically that Random-Walk Perturbations with Uniform Exploration Algorithm has the same regret degree as the existing optimal algorithm,and is superior to it in reallocation times.At the same time,experiments show that the proposed algorithm achieves better results compared with existing algorithms in general when fighting against various types of attackers.