Computational Intelligence

Course Objective

1. Knowledge and understanding of:
- Optimization techniques: Derivative-Free Optimization (Hill Climbing,
Local Search), Gradient-based Optimization (Gradient Descent and
Stochastic Gradient Descent)
- Evolutionary Algorithms
- Neural Networks (Fully-Connected Networks, Convolutional Networks)
- Sampling Methods: Metropolis-Hastings Algorithm, Simulated Annealing
- Reinforcement Learning: Q-learning
- Neuroevolution: Neural Architecture Search, NEAT.

2. Applying knowledge and understanding:
- How before mentioned optimization algorithms work and where to use
them.
- How to formulate an evolutionary algorithm for a specific problem.
- How to analyze an optimization algorithm.
- What neural network fits best for given problem.

3. Making judgments:
- What optimization algorithm to use for given optimization problem
(either Derivative-Free method or Gradient-based method).
- What layers of a neural network are suitable in a given problem.

4. Communication skills
- Presenting an analysis in a written form (a short report) for each
assignment.

Course Content

In the course Computational Intelligence, we will focus mainly on
computational aspects of Artificial Intelligence, namely, optimization
algorithms for solving learning problems. Specifically, we will consider
problems that cannot be solved using information about gradient due to
their combinatorial character or complexity of the objective function
(e.g., non-differentiability, blackbox objective function). These
problems pop up in computer science and AI, such as, identification of
biological systems, task scheduling on chips, robotics, finding optimal
architecture of neural networks. For this purpose, we will introduce
different classes of algorithms that can be used to tackle these
problems, namely, hill climbing and local search, and evolutionary
algorithms. Additionally, we explain sampling methods (Markov Chain
Monte Carlo) and population-based sampling methods, and indicate how
they are linked to evolutionary algorithms. In the second part of the
course, we will discuss neural networks as current state-of-the-art
modeling paradigm. We will present basic components of deep learning,
such as, different layers (e.g., linear layers, convolutional layers,
pooling layers, recurrent layers), non-linear activation functions
(e.g., sigmoid, ReLU), and how to use them for specific problems. At the
end of the course, we will touch upon alternative approaches to learning
using Reinforcement Learning and Bayesian Optimization. We will conclude
the course with a recently revived field of neuroevolution that aims for
utilizing evolutionary algorithms in training neural networks.

Additional Information Teaching Methods

Lectures and practical assignments.

Method of Assessment

The final grade is calculated based on the final exam (50 points) and 5 individual practical assignments (10 points each, 50 points in total). To pass the course, students are required to obtain at least 25 points from the final exam, and 55 points in total including all the points from the exam and assignments.

The exam can be retaken (a resit).

Solutions to practical assignments must be provided within given

deadlines. There is no resit option for practical assignments.

Literature

The literature will be made available on Canvas.

Additional Information Target Audience

Bachelor Artificial Intelligence (track Intelligent Systems)

Recommended background knowledge

Calculus
Algebra
Statistics
Programming in Python