Chapter 5
Question 1: a decision tree for the following Boolean functions is provided:(A branch B) Branch (C Branch D)
Question 2: Determine whether to have a picnic Based on the weather conditions. There is a collection that indicates whether to have a picnic based on certain conditions. Next to this question is a decision tree:
()Circled Table OriginalTrueHoweverFalseInstance
(B)OriginallyFalseHoweverTrueInstance
(C)If a leaf node is converted into an attribute set, which leaf node should be converted?
(I) Raining = T (ii) sunny = T (iii) sunny = f
(D)Which of the newly added attributes can provide more information? (Log_2 1/3 ≈ 1.6 ))
Note: This question is wrong. The set does not match the actual situation. I do not know whether the decision tree is used or the set prevails. The set is ambiguous.
A: (This question is purely a mathematical question. I will never take the test. The teacher once thought about changing this question. I did not change it at last, but it does not affect the study of AI, I didn't find the right one in my homework. It was a little too late for me to push it, so we just ignored it)
Question 4: Given7Three-dimensional data, requiring each cluster to contain at least one data and not limit the number of clusters after clustering, what is the possible clustering scheme?
Question 5: What is the essence of supervised learning computing? Which two key sub-questions are involved? Describe the solutions you know as much as possible for each subproblem.
A: The essence of supervised learning is the objective function for learning to fit actual problems. Two of the key problems are the optimization goal and method, the optimization objectives can be the least square, the maximum information entropy, and the minimum description length. The optimization methods mainly include the fastest descent, Newton iteration, evolutionary computing, and group intelligence.
Chapter 6
Question 1: briefly describe the main differences between behavioral intelligence and symbolic and connectionist.
A: In order to respond to external stimuli, The sympositism and connection doctrine hold that the intelligent system must be able to express the external environment and the goal of solving the problem, and then make reasoning on this basis to determine the action to be taken. It is usually represented in the "Perception-representation-inference-behavior" mode. Behavior intelligence is represented as a "Perception-behavior" model. The agent directly responds to the perceived external stimulus, and constantly adapt to the external environment in the "Perception-behavior" process.
Question 2: Trial in Reinforcement LearningQ-The learning method solves the following second-order fan ta problem:
It also describes the fundamental differences between the solution concept and the Graph Search Method in the problem solving.
(Note: you do not need to describe the learning process and give it directly.Q-Value, and an overview of its solution Principles)
A: In the form of a directed graph (each node is in various states and has a directed edge representing the State Transition), the two-layer tower chart is shown. When the intelligent system starts, the reward value for the target State is 1000, and the remaining State is 0.
Assume that the search starts from a certain State, enters the state that you think is correct with a 80% probability, and enters another State with a 20% probability (if any ), update your own Q value based on the Q value of the next state. This will continue to traverse the entire state and allow the SMART system to continuously learn and optimize its Q value during the traversal process, at last, the Q of each status tends to be stable.
After learning through Q-learning methods, the behavior intelligence system is adaptive to the environment, which is the fundamental difference from graph search methods. In the Graph Search Method, each search step needs to evaluate and decide the next step based on the search policy. In the Q learning method, once the intelligent body performs adaptive learning on the environment, each status has its own Q value, and can quickly find the target status regardless of the status.
Q:
Q-learningAlgorithmThe fundamental difference between graph search algorithms is that, no matter whether heuristic functions are used or not, the search process is completely blind (heuristic information is empirical and does not necessarily guide to the optimal solution ), the Q-function value of the Q-learning algorithm can accurately evaluate the action.