Machine Discovery and Machine Learning Based on Adaptation and Evolution

两年来的研究成果总结如下：(1)进化计算中的研究成果1)交叉算子在遗传算法中的作用分析我们对交叉算子在遗传算法中搜索更好解的作用进行了数学分析，并给出了“欺骗边界定理”。我们提出了完备性、健全性、非冗余性和特征保存性四个评价标准来评价模型公式。3)子游交换交叉作为字符保持操作的提出及应用在排序问题领域提出了一种字符保持操作子游交换交叉。4)多目标优化问题遗传算法的提出本文建立了一种遗传算法生成pareto最优集的方法，该方法是多目标优化问题的理性解。(2)强化学习框架和分类的提出我们定义了一个透明的强化学习框架，并根据环境的类别和方法的方向对现有的研究进行了分类。2)面向开发的强化学习合理性分析我们对利润分享方法中共享函数的合理性进行了数学分析，给出了两类合理性定理。为了有效识别未知环境，提出了一种面向探索的行动选择策略——k-确定探索方法。4)基于期望奖励爬坡的强化学习方法的提出我们提出了一种增量强化学习方法，它沿着期望奖励的梯度进行爬坡。5)感知混叠下学习方法的提出提出了一种预测模型来预测不完全欺骗感知下的状态转移，并开发了一种学习方法来构建一个试错预测模型。少

Research results through two years are summarized as follows ;(1) Research results in evolutionary computation1) An analysis on the role of crossover operators in genetic algorithmsWe made a mathematical analysis on the effect of crossover operators to search better solutions with genetic algorithms, and showed the "deceptive boundary theorem."2) A proposal of evaluation criteria on a coding-crossoverWe proposed four criteria of completeness, soundness, non-redundancy and character preservingness to evaluate a model formulation.3) A proposal and applications of the subtour exchange crossover as a character preserving operationWe proposed a character preserving operation named the subtour exchange crossover in the domain of ordering problems.4) A proposal of genetic algorithms for multiobjective optimization problemsWe established a methodology for genetic algorithms to generate the pareto optimal set, which is the rational solution for a multiobjective optimization.(2) Research results … More in adaptive computation1) A proposal of a framework and a categorization for reinforcement learningWe defined a transparent framework for reinforcement learning, and classified existing researches according to the class of the environment and the orientation of the approach.2) An analysis on the rationality of the exploitation oriented reinforcement learningWe made a mathematical analysis on the rationality of the sharing functions in the profit sharing method, and showed two kinds of rationality theorem.3) A proposal and an extention of the k-certain exploration methodWe proposed a exploration oriented action selection strategy named the k-certain exploration method to efficiently identify an unknown environment.4) A proposal of a reinforcement learning method based on a hillclimbing of the expected rewardsWe proposed an incremental reinforcement learning method, which performs a hillclimbing along the gradient of the expected rewards.5) A proposal of a learning method under perceptual aliasingWe proposed a prediction model to foresee the state transition under incomplete deceptive perception, and developed a learning method to construct a prediction model with trial and error. Less