Artificial Intelligence Assignment Sample

Q1.Difference between

1. Supervised and unsupervised learning

Supervised learning algorithm

Supervised as well as unsupervised learning are algorithms are important machine learning algorithms. Each technique is used with different scenarios.

Supervised machine learning is a technique in which models established through labelled data. In supervised machined algorithms, trained model map the input variables or features to output variable or label (Florensa, 2019).

Y=f(x)

‘Y’ is class label. The function is known as trained model. The testing data or new data will feed into model or function to predict the class label. The model will be trained with data or features as well as its class label. Supervised learning algorithms need supervision for training the model. The following two types of problem use supervised machine learning algorithms such as

• Classification
• Regression

Example of supervised learning algorithm

The dataset has images of different fruits. In supervised machine learning algorithm, the model is trained through the features or characteristic of different fruits such as shape, colour, and taste to classify fruits. The trained model will predict the fruit name for the test data.

Classification

Classification predicts the class label for unforeseen input data. When the classification predicts two distinct class label is known as binary classification. When the classification algorithms predict more than one class label then it is called as multiclass classification.

Regression

Regression algorithm predicts the value of the unforeseen data. for example, regression algorithm can predict the value of  house in next 5 years.

Unsupervised machine learning algorithm

Unsupervised machine learning algorithm do not train model from the predefined class label. The patterns as well as structure are identified through the input data. It does not have any supervision. The grouping of data or input will be done based on similar patterns as well as structure. The learning problems such as clustering as well as association belong to clustering technique.

Example for unsupervised learning algorithm

The dataset has images of fruits. The data of dataset does not have any label with it. The data set has only features or characteristics. The images are grouped based on the same features or characteristics.  The model will be trained based on those data. The grouping of test data will be done through the trained model.     

2. Complete and incomplete search

Complete search

Complete as well as incomplete search are two different techniques to solve combinatorial problems. Most of C-Constraint P-Programming (CP) as well as M-Mixed I-Integer P-Programming (MIP) are belong to complete.  Complete search algorithms are having desirable properties. Complete algorithms are complete on feasibility problem (SAT) that offers either solution or prove that there is no solution exists for the problem. Search algorithm on optimisation problem (job shop scheduling) that offers either optimum solution or it proves that there is no optimum solution exists for the problem. The algorithms do not have run time parameters. Various complete search algorithms exists in AI such as B-best F-first S-search (BFS), breadth first search as well as iterative deepening. The solvers such as MIP and CP belong to specific type of complete search such as D-depth F-first S-search (DFS). DFS algorithm is exhaustive. It finds all possibilities except relaxation constraint propagation or some form of reasoning. DFS has more advantage than breadth first search for consuming less memory.

Incomplete search

Metaheuristic algorithms are belonging to incomplete and redundant. The algorithms are failing to return solution for satisfiable problems.  The algorithms are also fail to produce optimal solution for optimization problems. Metaheuristics algorithms are designed for escaping from local minima as well as having the P-probabilistic A-approximate C-completeness (PAC) property. The probability for finding solution is 1 with the search time of infinity. One of the ways for periodically restarting search from random state such state is known as non-zero probability to get an optimal solution or being close to the optimal solution.  Metaheuristics algorithms cannot prove problems that are unsatisfiable. Local search algorithms such as C-complete L-local S-search-CLS as well as P-partial O-order D-dynamic B-backtracking- PODB rove both satisfiability and unsatisfiability.    The main aim of these problems is to improve the performance of satisfiable problem and not unsatisfiable problem.  Metaheuristics technique generally applied for unconstrained optimization problem, but it can also handle constraints problem also. The theorem No free lunch proves that no algorithm is better than each other.

3. Thinking humanly and thinking rationally

Thinking humanly means thinking as humans or persons. Cognitive science deals about the topic thinking humanly. The field is trying to make devices to think like humans. It models human thinking process (Crowder,2020).

Thinking rationally means model thinking as logical process. Symbolic logics are used to draw conclusions from processes. 

Q2. Answer the following

1. State the types of machine learning algorithms

Machine learning algorithms are classified into the following types such as

• Supervised
• Unsupervised
• Semi supervised
• Reinforcement

Supervised learning algorithm

Supervised machine learning algorithms are known as predictive model

• Nearest neighbour
• Naïve Bayes
• Decision tree
• SVM
• N-Neural N-Networks-NN

Unsupervised machine learning algorithms

The model is known as descriptive model. The following algorithms belong to unsupervised learning algorithms such as

• Association rule
• K-means clustering

Semi supervised learning algorithm

Semi supervised learning algorithm reside in between supervised and unsupervised algorithm. Supervised algorithm predicts the class label of input data.  Unsupervised learning descripts the data into group based on its characteristics or features. The algorithm will be applied to train model from both labelled as well as unlabelled data.

Reinforcement learning

It is type of machine learning. It deals about suitable decisions for maximizing reward in particular situations. It uses various machines as well as software for finding best possible path or behaviours according to specific situation.  Reinforcement learning do not have any training data. It learns through its experience. Example for such learning is te problem with robot, fire as well as diamond. The robot or gent has to reach reward such as diamond by crossing hurdles or fires.   The robot tries each path and gets experience as well as finally reaches reward from the gained experience.

Algorithms

• Q-L-Learning- QL
• T-Temporal D-Difference -TD
• D-Deep A-Adversarial N-Networks-DAN

2. State the basic operation of genetic algorithms

G-Genetic A-algorithms-GA are A-adaptive H-heuristics S-search A-algorithms-AHSA that belong to evolutionary algorithms. It is based on the key concepts such as genetics and natural selection. The algorithms are used to provide high quality solutions for search problems as well as optimization problem. It simulated the natural selection process. The concept of natural selection process deals about the species that adopts the changing environment can able to survive and reproduce as well as move to next generation. It simulates survival of fittest. Each generation includes individual populations and each individual demonstrates point in search space as well as possible solution. String of integer/bits/floats/character represents each individual (Steven,2019).

Search space of genetic algorithm is depicted below:

Operators and operations of GA

• Selection operation

The operation is giving preference to individuals that have good fitness score and allowing them to pass into genes for successive generations. 

• Crossover operation

The operation performs mating between individuals. Selection operator used to select individuals.  Cross over operator selects the cross over sites on individuals as well as selecting genes of these cross over sites are exchanging to create new individual.

Example of crossover operation is depicted below:

Mutation operation

Mutation is the task of inserting random genes in new genes for maintaining the diversity in population for avoiding premature convergence.

The mutation operation is illustrated below:

III- For an intelligent agent that can wipe the windshields when needed, state the components required for each of: Perception, Sensors, Effectors, and Actions

Windshield wiper agent

The Goal of an agent is to keep windshield clean as well as maintain the visibility

Perception: Dirty, raining

Sensors : camera

Effectors : wipers (right, left, back)

Actions : slow, medium, fast, off

Environment : weather, inner city and highways

b. impact of over fitting on accuracy

The model is selected based on its accuracy on training data rather than on its unforeseen test data. When such model is applied on test data, the accuracy will get reduced due to its specialized characteristic of the model. It indicates that, the model is not generalized. This situation is called as over fitting. Over fitting occurs when the model learns the detailed specification of training data such as noise. It will negatively impact the performance of machine learning algorithms. The over fitted model is learned from the fluctuation samples as well as noises (Montoya, 2020).  While applying the new test data on the model then the learned concept cannot applicable. It will lead to poor accuracy.

Machine learning algorithm performance is getting affected due to over fitting or under fitting.  

b. Is this kind of search complete? Discuss

Yes. This kind of search is complete. DFS is recursive algorithm that uses the concept of backtracking. It uses exhaustive searches as well as nodes move ahead when possible otherwise backtracking. The nodes are moving ahead when the nodes are exists along the path otherwise it moves back on the same path to find nodes for traversing. For the given problem, DFS provides the solution such as shortest path from Fargaras to Cratiova.

References

• Steven Prestwich (2008). The Relation Between Complete and Incomplete Search, Springer
• Crowder, J. A., Carbone, J., & Friess, S. (2020). Implicit Learning in Artificial Intelligence. In Artificial Psychology (pp. 139-147). Springer, Cham.
• Crowder, J. A., Carbone, J., & Friess, S. (2020). Systems-Level Thinking for Artificial Intelligent Systems. In Artificial Psychology (pp. 15-27). Springer, Cham.
• Montoya, J. H., Winther, K., Flores, R. A., Bligaard, T., Hummelshøj, J. S., & Aykol, M. (2020). Autonomous intelligent agents for accelerated materials discovery.
• Florensa, C., Degrave, J., Heess, N., Springenberg, J. T., & Riedmiller, M. (2019). Self-supervised learning of image embedding for continuous control. arXiv preprint arXiv:1901.00943.