Univ.

Exam Roll o Definition: Uses a mix of labeled and unlabeled data for
No.
training.
o Examples: Learning with small labeled datasets
Dr. K. N. Modi Institute of Engineering and Technology
combined with large unlabeled datasets.
MODINAGAR
4. Reinforcement Learning
o Definition: The model learns by interacting with the
ODD Semester Academic Session 2024-25
environment to maximize rewards.
Pre-University Test o Examples: Game-playing AI, Robotics.
Branch: o Algorithms: Q-Learning, Deep Q-Networks, Policy
Program: B.Tech CSE Sem. Vth Gradient Methods.
Subject Name: Machine Learning 5. Self-Supervised Learning
Subject Code: BCS055 Techniques
o Definition: A form of unsupervised learning where the
Total Marks:
Time: 3 Hr. 70 model generates its own labels.
o Examples: Contrastive learning, BERT for NLP.
6. Evolutionary Algorithms
Note: Attempt all Sections. If require any missing data; then choose o Definition: Uses optimization techniques inspired by
suitably.
natural evolution.
SECTION – A o Examples: Genetic Algorithms, Particle Swarm
Optimization.
2x7=
1) Attempt all questions in brief. 14 b) What are the steps used for making decision tree ?
a) Describe different types of machine learning algorithm.
Ans: Steps for Making a Decision Tree
Ans: Different Types of Machine Learning Algorithms
1. Select the Root Node: Start by choosing the best attribute using a
1. Supervised Learning selection criterion (e.g., Gini Index, Information Gain).
o Definition: The model learns from labeled data, mapping input to 2. Split the Dataset: Partition the dataset based on the selected attribute.
output. 3. Create Subtrees: Repeat the process recursively for each subset.
o Examples: Regression, Classification. 4. Stopping Criteria: Stop when all data points in a node belong to the
o Algorithms: Linear Regression, Logistic Regression, Support same class, or when further splitting adds no significant improvement.
Vector Machines, Neural Networks. 5. Pruning (Optional): Simplify the tree to avoid overfitting by removing
2. Unsupervised Learning branches with low importance.
o Definition: The model works with unlabeled data to find structure
or patterns.
o Examples: Clustering, Dimensionality Reduction.
o Algorithms: K-Means, DBSCAN, PCA (Principal Component c) Describe the difficulties faced in estimating the accuracy of hypotheses.
Analysis).
3. Semi-Supervised Learning Ans: Difficulties in Estimating the Accuracy of Hypotheses
 1. Bias-Variance Tradeoff: Balancing bias and variance is challenging, as too o VC Dimension: Measures the capacity of a hypothesis class.
much focus on training accuracy may lead to overfitting. o Sample Complexity: Minimum number of training samples
2. Insufficient Data: Limited or non-representative data can distort accuracy required for accurate learning.
estimates. o Generalization Bound: Theoretical limits on model
3. Imbalanced Datasets: Skewed class distributions make accuracy metrics performance for unseen data.
misleading.
4. Cross-Validation Errors: Improper cross-validation techniques can yield f) How to avoid overfiting in decision tree model ?
unreliable estimates. Ans: Avoiding Overfitting in Decision Tree Models
5. Overfitting: High accuracy on the training set may not generalize well to
unseen data. 1. Pruning: Remove branches that have little impact on overall accuracy.
2. Restrict Tree Depth: Limit the depth to reduce complexity.
3. Minimum Samples: Set a threshold for the minimum number of
samples required for splitting.
d) Explain different phases of genetic algorithm. 4. Cross-Validation: Use techniques like k-fold cross-validation to test
the model’s robustness.
5. Feature Selection: Use only the most relevant features.
6. Regularization: Apply penalties for overly complex trees.
Ans: Phases of Genetic Algorithm
g) What are the applications of machine learning ?
1. Initialization: Generate a population of potential solutions (chromosomes)
randomly. Ans: Applications of Machine Learning
2. Selection: Choose the fittest individuals based on a fitness function.
3. Crossover: Combine pairs of individuals to produce offspring by 1. Healthcare: Disease prediction, medical imaging analysis, drug
exchanging genetic material. discovery.
4. Mutation: Introduce random changes to maintain diversity and explore 2. Finance: Fraud detection, credit scoring, algorithmic trading.
new solutions. 3. Retail: Recommendation systems, inventory management, customer
5. Evaluation: Compute the fitness of new offspring. segmentation.
6. Termination: Stop when a termination criterion (e.g., maximum 4. Autonomous Vehicles: Object detection, path planning, and driving
generations, optimal fitness) is met. automation.
5. Natural Language Processing: Chatbots, sentiment analysis,
e) Write short note on computational learning theory. translation.
6. Manufacturing: Predictive maintenance, quality control, robotics.
Ans: Computational Learning Theory 7. Agriculture: Crop monitoring, yield prediction, pest control.
8. Gaming: Game AI, player behavior analysis.
 Definition: A subfield of AI and ML that focuses on quantifying the
computational complexity and feasibility of learning algorithms.
 Key Concepts:
SECTION – B
o PAC Learning (Probably Approximately Correct): Framework for
assessing learning feasibility. 2) Attempt any three of the following: 7x3=
 21 4. Reinforcement Learning
a) What are the advantages and disadvantages of different types of
machine learning algorithm ?  Advantages:
o Ideal for decision-making problems.
Ans: Supervised Learning o Learns from interactions and feedback.
o Handles dynamic environments well.
 Advantages:  Disadvantages:
o Produces accurate predictions for labeled data. o Requires significant computation and time.
o Easy to understand and implement for simple problems. o Difficult to design reward functions.
o Useful in regression and classification tasks. o May converge to suboptimal policies.
 Disadvantages:
o Requires large amounts of labeled data.
o May overfit if the model is too complex.
b)Write short note on Artificial Neural Network (ANN).
o Struggles with high-dimensional data without feature selection.
Ans: An Artificial Neural Network (ANN) is a computational model inspired
2. Unsupervised Learning by the human brain's structure and functionality. It consists of layers of
interconnected nodes (neurons):
 Advantages:
o Works with unlabeled data. 1. Structure:
o Identifies hidden patterns or clusters in the data. o Input Layer: Receives input features.
o Reduces dimensionality and noise. o Hidden Layers: Process inputs using weights, biases, and
 Disadvantages: activation functions.
o Difficult to evaluate results. o Output Layer: Produces the final prediction or classification.
o May produce meaningless clusters without proper tuning. 2. Key Features:
o Sensitive to initialization and scaling of data. o Activation Functions: Introduce non-linear transformations.
Examples: ReLU, Sigmoid, Tanh.
3. Semi-Supervised Learning o Learning: Uses backpropagation to adjust weights to minimize
error.
 Advantages: o Types: Feedforward Neural Networks, Convolutional Neural
o Reduces labeling costs. Networks (CNNs), Recurrent Neural Networks (RNNs).
o Combines the benefits of supervised and unsupervised learning.
3. Applications:
o Image and speech recognition, natural language processing,
 Disadvantages:
and robotics.
o Performance depends on the quality of labeled and unlabeled
data.
o Algorithm complexity increases.
c) Describe the difficulties faced in estimating the accuracy of hypotheses.

Ans: Difficulties Faced in Estimating the Accuracy of Hypotheses

 1. Insufficient Data: o Choice of metric (e.g., Euclidean, Manhattan) affects
o Small datasets may not represent real-world scenarios adequately. algorithm performance.
2. Overfitting: 5. Noise Sensitivity:
o Models fit too closely to training data, reducing generalization to o Algorithms may be sensitive to noisy or irrelevant features in
unseen data. the dataset.
3. Underfitting: 6. Memory Usage:
o Models fail to capture underlying patterns due to excessive o Requires storing all training instances, which can become
simplicity. expensive for large datasets.
4. Class Imbalance:
o Imbalanced datasets lead to biased predictions toward majority
classes.
5. Evaluation Metrics: e) Explain different phases of genetic algorithm.
o Inappropriate or single metrics (e.g., accuracy) may not capture
true performance, especially for imbalanced data. Ans: Phases of Genetic Algorithm
6. Noise in Data:
o Presence of irrelevant or mislabeled data distorts accuracy 1. Initialization:
estimation. o Generate an initial population of solutions randomly or based
7. Cross-Validation Issues: on heuristics.
o Improper splitting can lead to biased estimates of model 2. Selection:
performance. o Choose individuals based on a fitness function to participate in
reproduction. Common methods: Roulette Wheel, Tournament
d) What are the performance dimensions used for instance Selection.
based learning algorithm ? 3. Crossover:
o Combine genetic material from two parents to create
Ans: Performance Dimensions for Instance-Based Learning Algorithms offspring. Methods: Single-point, Multi-point, Uniform
crossover.
4. Mutation:
Instance-based learning algorithms (e.g., K-Nearest Neighbors) rely on comparing
o Introduce random changes to offspring to maintain diversity
new instances with stored training data. Performance dimensions include:
and explore new solutions.
5. Evaluation:
1. Accuracy:
o Assess the fitness of new individuals to determine their
o Correctness of predictions compared to actual labels.
quality.
2. Efficiency:
6. Replacement:
o Time and computational resources required for training and
o Decide which individuals from the old population will be
prediction.
replaced by new offspring.
3. Scalability:
7. Termination:
o Ability to handle large datasets without significant performance
o Stop when a termination criterion is met (e.g., maximum
degradation.
generations, desired fitness level).
4. Distance Metric Selection:
 3. Generalization:
o Helps design models that perform well on unseen data by
SECTION – C focusing on relevant tasks and metrics.
4. Comparison:
7x1 o Facilitates benchmarking between different algorithms and
3) Attempt any one part of the following: =7
models.
a) Describe well defined learning problems role's in machine learning
b) What is learning ? Explain the important components of learning.
Ans: A well-defined learning problem provides a clear framework for
designing, implementing, and evaluating machine learning models. Such ANs: Learning is the process by which a system improves its performance on a
problems specify the task, performance metric, and the experience that specific task by gaining experience and adapting its behavior based on that
the model uses for improvement. experience. In the context of machine learning, it involves deriving patterns or
making predictions from data.
Components of a Well-Defined Learning Problem:
Important Components of Learning:
1. Task:
o Describes the specific problem the model aims to solve. 1. Data (Experience):
o Example: Predicting house prices, classifying emails as spam, or o The information the system uses to learn.
recognizing handwritten digits. o Examples: Training datasets, sensor readings, or user
interactions.
2. Performance Metric:
o Measures how well the model performs the task. 2. Model (Representation):
o Examples: Accuracy, precision, recall, F1-score, mean squared o The mathematical framework or algorithm used to represent
error. patterns in data.
o Examples: Linear regression, decision trees, neural networks.
3. Experience:
o Refers to the data or interactions the model uses for training and 3. Objective (Task):
improvement. o The specific goal of the learning process.
o Examples: Historical datasets, user interactions, simulated o Examples: Classification, regression, clustering.
environments.
4. Performance Metric:
Role of Well-Defined Problems in Machine Learning: o A criterion to evaluate how well the model is learning and
performing the task.
1. Clarity in Objectives: o Examples: Accuracy, precision, recall, mean absolute error.
o Ensures the problem is framed correctly for meaningful solutions.
2. Evaluation: 5. Learning Algorithm:
o Enables consistent measurement of the model's success or o The method used to train the model, adjusting its parameters
failure. based on the data.
 o Examples: Gradient descent, backpropagation, reinforcement o Description: Neurons can form loops, allowing information to
learning algorithms. persist and be reused.
o Architecture:
6. Generalization:  Recurrent Neural Network (RNN): Each neuron
o The model’s ability to apply what it has learned to unseen data or connects to itself or other neurons in the same layer.
tasks.  Variants: LSTMs (Long Short-Term Memory) and
GRUs (Gated Recurrent Units) for handling long-term
7. Feedback (Optional): dependencies.
o Information that helps refine learning. o Applications: Time series prediction, language modeling,
o Example: Rewards in reinforcement learning, corrections in supervised speech recognition.
learning. o Advantages: Memory of past inputs, good for sequential data.
o Disadvantages: Vanishing/exploding gradient problems.

7x1 3. Convolutional Connections:

4) Attempt any one part of the following: =7
o Description: Neurons are connected locally, focusing on
a) Explain different types of neuron connection with architecture.
specific regions of the input data.
o Architecture:
Ans: Neurons in artificial neural networks are connected in various ways,  Convolutional Neural Network (CNN): Includes
forming different architectures. These connections define how information convolutional layers, pooling layers, and fully
flows through the network and determine its capability to solve specific connected layers.
problems. o Applications: Image and video processing, object recognition.
o Advantages: Efficient for spatial data, translation invariance.
Types of Neuron Connections o Disadvantages: Computationally expensive for large-scale
data.
1. Feedforward Connections:
o Description: Information flows in one direction, from input to
4. Fully Connected Layers:
output, without loops. o Description: Every neuron in one layer connects to every
o Architecture: neuron in the subsequent layer.
 Single-Layer Perceptron: Consists of one input layer o Architecture:
directly connected to an output layer.  Found in the final stages of most architectures for
 Multi-Layer Perceptron (MLP): Includes one or more classification or regression tasks.
hidden layers between input and output layers. o Applications: General-purpose tasks.
o Applications: Classification, regression. o Advantages: Maximizes information flow between layers.
o Advantages: Simplicity, good for structured data. o Disadvantages: High computational cost and risk of
o Disadvantages: Limited in handling sequential or temporal data. overfitting.

2. Recurrent Connections: b) Explain adaline network with its architecture.

Ans: Adaline (Adaptive Linear Neuron) is an early neural network model 1. Define the Population:
that uses a single-layer architecture. It is similar to the perceptron but differs o Clearly identify the entire dataset or population from which
in its learning mechanism, employing a linear activation function and samples will be drawn.
gradient descent for weight updates. o Example: Customers of an e-commerce platform.
2. Determine the Sampling Objective:
Architecture: o Decide why sampling is necessary (e.g., reducing
computational costs, exploratory analysis).
3. Select the Sampling Technique:
1. Components:
o Choose an appropriate method based on the data and
o Input Layer: Accepts inputs (features) x1,x2,…,xnx_1, x_2, \ldots, x_nx1
,x2,…,xn. objectives:
o Weights: Each input xix_ixi is associated with a weight wiw_iwi.  Random Sampling: Each element has an equal chance
o Summation Function: Computes a weighted sum y=∑i=1nwixi+by = \ of being selected.
sum_{i=1}^n w_i x_i + by=∑i=1nwixi+b, where bbb is the bias.  Stratified Sampling: Data is divided into strata, and
o Output Layer: Produces the final output, which is compared with the samples are taken proportionally.
target value.  Systematic Sampling: Select every nnn-th element
from the population.
 Cluster Sampling: Randomly select clusters instead of
2. Activation Function:
individual elements.
o Uses a linear activation function.
4. Determine Sample Size:
o Unlike the perceptron, it does not use a step function but minimizes the
o Decide how many samples are needed for statistical validity or
error using the least mean squares (LMS) rule.
desired accuracy.
5. Collect the Sample:
Working of Adaline:
o Implement the chosen sampling method and extract the subset.
6. Validate the Sample:
1. Compute Output: Calculate the weighted sum of inputs.
o Ensure the sample represents the population accurately and
2. Compare with Target: Determine the error by subtracting the actual output from
the target. check for biases.
3. Weight Update: Adjust weights using the rule: wi=wi+η⋅(t−y)⋅xiw_i = w_i + \eta \ 7. Analyze the Sample:
cdot (t - y) \cdot x_iwi=wi+η⋅(t−y)⋅xi where η\etaη is the learning rate, ttt is the o Use the sampled data for statistical analysis, model training, or
target, and yyy is the output. hypothesis testing.

b) Differentiate between unsupervised and reinforcement learning.

Ans:
7x1 Aspect Unsupervised Learning Reinforcement Learning
5) Attempt any one part of the following: =7
a) What are the steps used in the process of sampling ?
Finds hidden patterns or Learns optimal actions
Definition structures in unlabeled through interaction with
Ans: Sampling is the process of selecting a subset of data from a larger dataset or data. an environment.
population to analyze or train models. The steps typically involved are: Feedback in the form of
Input Data Unlabeled data.
rewards or penalties.
 Aspect Unsupervised Learning Reinforcement Learning o Obtain the dataset containing features and corresponding labels
(for classification) or continuous values (for regression).
Groups (clusters),
Policy that maps states to
Output dimensionality reduction,
actions. 2. Choose the Number of Neighbors (K):
or associations.
o Decide on the number of nearest neighbors (KKK) to consider for
Understand the data Maximize cumulative making predictions.
Goal
structure or relationships.
rewards over time. o A small KKK may lead to noise sensitivity, while a large KKK might
Learning Based on similarity, Trial-and-error-based smooth out local patterns.
Process density, or association.learning from feedback.
Q-Learning, Deep Q- 3. Measure Distance:
Examples of K-Means, DBSCAN, PCA, o Compute the distance between the new data point and all other
Networks (DQN), Policy
Algorithms Hierarchical Clustering. points in the training data.
Gradient.
o Common distance metrics:
Market segmentation,  Euclidean Distance: d=∑i=1n(xi−yi)2d = \sqrt{\
Game AI, robotics,
Applications anomaly detection, image sum_{i=1}^{n} (x_i - y_i)^2}d=∑i=1n(xi−yi)2
autonomous vehicles.
compression.  Manhattan Distance: d=∑i=1n∣xi−yi∣d = \sum_{i=1}^{n} |x_i
No dependency on Considers sequences and - y_i|d=∑i=1n∣xi−yi∣
Temporal Aspect
sequence or time. future states.
Dependency on No interaction with an Requires interaction with 4. Identify K Nearest Neighbors:
o Sort the distances and select the KKK closest points to the new data
Environment external environment. a dynamic environment.
point.
Clusters, reduced Evaluated using
Evaluation dimensions, or associations cumulative reward or 5. Make Predictions:
are evaluated qualitatively. policy quality. o For Classification:
4o  Assign the class with the majority vote among the KKK
neighbors.
7x1 o For Regression:
6) Attempt any one part of the following: =7  Take the average (or weighted average) of the values of
a) Describe K-Nearest Neighbour algorithm with steps the KKK neighbors.

Ans. K-Nearest Neighbor (K-NN) is a simple, non-parametric, instance- 6. Evaluate the Model:
based machine learning algorithm used for classification and regression o Use metrics like accuracy (classification) or mean squared error
tasks. It predicts the output for a data point based on the outputs of its (regression) to evaluate performance.
closest neighbors in the feature space.
b) Briefly explain the inductive learning problem.
Steps in K-NN Algorithm:
Ans: Inductive Learning involves generalizing from specific examples
1. Load the Data: to form rules or hypotheses that can predict outcomes for new, unseen
examples. The main challenge is finding a hypothesis that correctly explains 7x1=
the training data and generalizes well to unseen data. 7) Attempt any one part of the following: 7
a) What is reinforcement learning? Explain passive reinforcement & active
Key Concepts in Inductive Learning: reinforcement learning ?

1. Objective: Ans: Reinforcement Learning (RL) is a machine learning paradigm

o Learn a function fff that maps input features XXX to output labels YYY where an agent learns to make decisions by interacting with an
using a given set of training examples (X,Y)(X, Y)(X,Y). environment. The agent receives feedback in the form of rewards or
penalties based on its actions and aims to maximize cumulative rewards
2. Generalization: over time.
o The ability of the learned hypothesis to perform well on unseen data.
Key Elements of Reinforcement Learning:
3. Hypothesis Space:
o The set of all possible hypotheses the algorithm considers. A smaller 1. Agent: The decision-making entity.
hypothesis space may oversimplify the problem, while a larger one can 2. Environment: The system with which the agent interacts.
lead to overfitting. 3. State (SSS): The current situation or configuration of the environment.
4. Action (AAA): Choices available to the agent at each state.
4. Bias-Variance Tradeoff: 5. Reward (RRR): Feedback signal from the environment indicating the success
o A balance is needed between a hypothesis that is too simple (high bias) or failure of the agent’s action.
and one that is too complex (high variance). 6. Policy (π\piπ): The strategy that the agent uses to decide actions.
7. Value Function (VVV): Estimates the future reward from a given state.
8. Q-Value (Q(s,a)Q(s, a)Q(s,a)): Estimates the reward of taking action aaa in
state sss.
Example of Inductive Learning:
Passive Reinforcement Learning:
 Scenario: Predicting whether an email is spam.
o Training Data: Labeled emails with features like subject keywords,  In passive reinforcement learning, the agent follows a fixed policy π\piπ and
sender information, etc. evaluates how good the policy is without trying to improve it.
o Goal: Generalize from labeled examples to predict whether new emails  The goal is to estimate the value function (Vπ(s)V^\pi(s)Vπ(s)) for the given
are spam or not. policy.

Challenges in Inductive Learning: Characteristics:

1. Noise in Data: Mislabeled or irrelevant data may mislead the learning algorithm. 1. Fixed Policy: The agent does not explore alternative actions.
2. Underfitting and Overfitting: Ensuring the learned hypothesis is neither too 2. Evaluation: Learns the value of states under the fixed policy.
simplistic nor too complex. 3. Learning Methods:
3. Computational Complexity: Searching through a large hypothesis space can be o Monte Carlo Methods: Estimate value functions by averaging
computationally expensive. rewards from complete episodes.
o Temporal Difference (TD) Learning: Updates the value function
using differences between consecutive estimates: V(s)←V(s)
 +α[R+γV(s′)−V(s)]V(s) \leftarrow V(s) + \alpha \left[ R + \gamma V(s') -  a′a'a′: Next action.
V(s) \right]V(s)←V(s)+α[R+γV(s′)−V(s)]  RRR: Immediate reward.
 α\alphaα: Learning rate.
Active Reinforcement Learning:  γ\gammaγ: Discount factor.

 In active reinforcement learning, the agent explores the environment and Steps in the Q-Learning Process:
improves its policy over time to maximize cumulative rewards.
 The agent learns an optimal policy π∗\pi^*π∗, which determines the best action 1. Initialize Q-Table:
to take in any state. o Create a table with all possible state-action pairs initialized to 0 or
small random values.
Characteristics:
2. Loop Until Convergence:
1. Policy Improvement: The agent dynamically updates its policy based on the o Repeat for each episode:
feedback. 1. Initialize State:
2. Exploration vs. Exploitation:  Start in an initial state sss.
o Exploration: Try new actions to discover better policies. 2. Choose Action:
o Exploitation: Use the current best-known policy to maximize rewards.  Use an action selection strategy (e.g., ϵ\epsilonϵ-
3. Learning Methods: greedy):
o Q-Learning: Learns the optimal action-value function (Q∗(s,a)Q^*(s,  With probability 1−ϵ1-\epsilon1−ϵ, select
a)Q∗(s,a)). the action with the highest Q(s,a)Q(s,
o Policy Gradient Methods: Optimize the policy directly using gradient- a)Q(s,a).
based methods.  With probability ϵ\epsilonϵ, select a
random action to explore.
b) Describe Q-learning algorithm processes 3. Take Action and Observe Reward:
 Perform the selected action aaa.
 Observe the immediate reward RRR and the next
Ans: Q-Learning is a model-free reinforcement learning algorithm that
state s′s's′.
enables an agent to learn the optimal action-value function Q∗(s,a)Q^*(s, 4. Update Q-Value:
a)Q∗(s,a) by interacting with the environment.  Update Q(s,a)Q(s, a)Q(s,a) using the Q-learning
update rule.
Key Concepts: 5. Transition:
 Set s←s′s \leftarrow s's←s′ for the next iteration.
1. Action-Value Function (Q(s,a)Q(s, a)Q(s,a)):
o Measures the expected cumulative reward of taking action aaa in state 3. Convergence:
sss, followed by the optimal policy. o Repeat the above steps for multiple episodes until Q(s,a)Q(s,
2. Bellman Equation: a)Q(s,a) stabilizes or reaches convergence.
o Q-Learning uses the Bellman optimality equation to iteratively update
Q(s,a)Q(s, a)Q(s,a): Q(s,a)←Q(s,a)+α[R+γmax⁡a′Q(s′,a′)−Q(s,a)]Q(s, a) \
leftarrow Q(s, a) + \alpha \left[ R + \gamma \max_{a'} Q(s', a') - Q(s, a) \
right]Q(s,a)←Q(s,a)+α[R+γa′maxQ(s′,a′)−Q(s,a)] where:
 s′s's′: Next state.
Advantages of Q-Learning:

 Model-Free: Does not require knowledge of the environment’s transition

probabilities.
 Optimal Policy: Converges to the optimal policy with sufficient exploration.

Challenges:

 Exploration vs. Exploitation Tradeoff: Balancing exploration and exploitation is

crucial.
 Scalability: Q-Learning struggles with large state or action spaces (addressed by
Deep Q-Learning).

Applications of Q-Learning:

 Robotics (e.g., motion control).

 Game AI (e.g., learning strategies in video games).
 Traffic signal optimization.
 Personalized recommendations