feat: parsed text and my solutions

Author: lmProgramming

lecture/2025/ang1_solved.png

@@ -0,0 +1,101 @@
+1. Can we effectively formulate every problem as a state space search problem?
+a) No, because we will have too many operators and too high a branching factor in the search graph.
+b) No, because not all problems allow us to define inverse transition operators between states.
+c) Yes, because complex operations can be represented as operators acting on a computational graph.
+d) Yes, because symbolic-arithmetic operations can be recursively represented in a computational graph.
+
+2. Is there a place for a deduction-based approach in contemporary artificial intelligence?
+a) No, because practically all contemporary AI systems are based on machine learning.
+b) Yes, deductive methods allow for the expression of verified knowledge, e.g., in the form of rules, also as sources of knowledge for generative language models (LLMs).
+c) Yes, because deductive methods are essential for data augmentation.
+d) Yes, because deductive methods are present in every supervised machine learning algorithm as an element of controlling the learning process.
+
+3. What is arc consistency in constraint satisfaction problems and why is it used?
+a) Arc consistency is the consistency of the path of connected variables in search and an element of partial solution representation.
+b) Arc consistency guarantees that for any two variables connected by a constraint, their domains contain values that satisfy that constraint.
+c) Arc consistency is a transformation of the problem, resulting in it consisting of n subproblems, in which the pairs are connected by a single arc.
+d) Arc consistency is a partial order of pairs: variable—heuristic, accelerating the solution finding.
+
+4. What is the partial order planning algorithm based on?
+a) Hierarchically dividing the problem into areas of local subplans that are next combined.
+b) The lack of distinction between actions that are generated for the same operators.
+c) Utilising only the necessary constraints in the representation of the plan.
+d) A partial order plan allows for conflicting sequences of actions, where conflicts are resolved later through additional searching.
+
+5. What are the true properties of the evaluation function in algorithms for playing logical games?
+a) The evaluation function must be consistent with the payoff function for terminal nodes and estimate it for all intermediate states.
+b) It must return 0 for the terminal node of the game tree and be consistent with the payoff function for the others.
+c) The evaluation function is a heuristic function for the cost of winning in one move.
+d) The evaluation function is any linear function of the set of features describing the game state.
+
+6. Why is it important for the heuristic to be admissible in state space search?
+a) An admissible heuristic does not hide the problem and returns values no higher than the actual cost.
+b) Because it allows for comparing values between nodes.
+c) Because only an admissible heuristic defines equivalence classes of nodes based on its values.
+d) Because it does not hide the path leading to a node that satisfies the goal test.
+
+7. When do positions in the game tree become stable (quiescent)?
+a) Stable (quiescent) positions are those for which there is a balance between attacking moves and moves that create threats.
+b) Positions in the game tree become stable (quiescent) when their subtrees become balanced after applying alpha-beta pruning.
+c) Stable (quiescent) positions are those for which the situation in the game does not correlate with significant changes in the value of the evaluation function over the next few moves.
+d) Stable (quiescent) positions are below the depth that defines the horizon problem in the game tree.
+
+8. How does constraint propagation affect the efficiency of a constraint problem-solving algorithm?
+a) It is an alternative approach to backtracking search based on heuristic forward searching.
+b) Constraint propagation shifts the search to the level of local operations within the variable domains.
+c) Constraint propagation dynamically reduces the effective size of the variable domains.
+d) It breaks the problem into independent subproblems of constraint solving within individual pairs of variables.
+
+9. Do planning algorithms differ from heuristic state space search?
+a) Planning algorithms are heuristic search algorithms based on a partial order graph heuristic.
+b) Planning algorithms utilise knowledge from the symbolic representation of state change operators.
+c) Planning algorithms do not build a complete transition graph from the initial state to the final state, but only a forest of partial graphs.
+d) Planning algorithms build the action graph deterministically and later apply heuristics to linearise the partial order graph into the final graph.
+
+10. Why does unsupervised machine learning work and to what extent?
+a) Unsupervised machine learning operates on the basis of transformation of some attributes into decision classes based on correlation analysis.
+b) Unsupervised machine learning explores similarities between objects to make property transfers or determine equivalence classes.
+c) Unsupervised machine learning relies on the augmentation of labelled data based on data clustering.
+d) Unsupervised machine learning only works when the results of data clustering are correlated with an accepted class hierarchy for the data.
+
+11. What is the order of assigning attributes to the nodes of the decision tree?
+a) Any order is acceptable as long as the attributes with the highest InfoGain value are chosen.
+b) A single path contains a group of attributes determined by a clustering algorithm and then subjected to unsupervised attribute selection.
+c) The assignment of attributes to nodes results from an unsupervised contextual assessment of cross-entropy.
+d) The order in each path of the tree (root - leaf) results from the reduction of uncertainty.
+
+12. What needs to be done to adapt decision trees, which are inherently symbolic, to operate on numerical values?
+a) It is necessary to quantize numerical attributes in a way statistically adjusted in terms of intervals for each attribute individually.
+b) Functions should be introduced that assign appropriate symbols to numerical values from the training data.
+c) Numerical values that may appear in the input data can simply be treated as individual symbols.
+d) To the tree nodes corresponding to numerical values, functions should be assigned that calculate the choice of the outgoing branch.
+
+13. How does training a neural network change it for specific tasks?
+a) The network remains the same, for example, the number of neurons and their types do not change; only the structure of the network is reorganized, including the positions of selected neurons, according to individual decision paths.
+b) As a result of training, activation paths related to the recognition of specific classes are created within the neural network during its use.
+c) Training the network causes the network parameters to adjust gradient-wise to a softmax probability distribution correlated with the given set of classes.
+d) As a result of training, the connections between neurons are adjusted, and some may even be suppressed, due to changes in weights and the distribution of values generated at the outputs.
+
+14. What is the significance of document segmentation for the functioning of the RAG system?
+a) Segmentation determines the detail of the response by defining the size of the text window as the basis for providing the answer.
+b) The size and methods of segmenting text (fragments of documents) can facilitate the precise determination of the basis for generating a response.
+c) Segmentation does not affect the precision of the responses given – they are generated by a large language model (LLM), but it is useful for processing efficiency.
+d) Document segmentation in the database facilitates the generation of concise responses of limited length.
+
+15. How does a neural network classify input data?
+a) Each layer of the network defines the projection of input signals onto the set of classes; and at the end, the output layer performs aggregation and voting according to the softmax scheme.
+b) The neural structures of a non-recurrent network form a forest of fuzzy classification trees built from subsets of neurons. 
+c) Activation functions shape the level of output signals, and in the output layer, the signals from the neurons are interpreted according to the assumed classification task.
+d) The neural network in the input layer selects attributes and processes their chosen subset into a stimulation vector, which then passes through the subsequent layers and influences the output neurons in a gradient-wise way.
+
+16. What can we use the k-nearest neighbors (k-NN) algorithm for?
+a) It allows for making assumptions about the selected properties of unknown objects based on the exploration of an assumed similarity measure.
+b) The k-NN algorithm can be used to evaluate the results of classification by a supervised algorithm.
+c) It enables the generation of a hierarchical structure of classes for the input data.
+d) It allows for discovering the optimal partitioning of input data and determining the number of classes – the parameter k – that are represented in that data.
+
+17. Why do we distinguish between the retrieval and re-ranking phases in Semantic Retrieval?
+a) This is due to the use of networks of varying depths and the orthogonal perspectives on comparing queries and documents.
+b) Because the matching of the query in the RAG system gradually transitions from the document level, through fragments, to sentences.
+c) The goal is to maximize the accuracy of the cut-off ranking of the retrieved documents while minimizing computational resource consumption.
+d) Both phases define different perspectives on comparing texts: sentence to sentence and word to word in the second case.