Methods of Decision-Making in Engineering

Active Learning and Incremental Learning

• Active learning involves actively choosing to label data that is informative, such as when many classifier candidates disagree.
• Incremental learning is a dynamic technique where input data is continuously used to extend an existing model's knowledge, such as when data becomes available gradually over time or exceeds memory limits.

Online Learning

• Online learning involves training a system incrementally by feeding it data instances sequentially, either individually or in mini-batches, which is processed in real-time to continuously update the model.

Explainability

• Explainability is crucial to understand what input to change to obtain a desired output and for debugging.
• Explanations can be global (entire data) or local (near data point), inherent or post-hoc (obtained through analysis, model-based for specific model type or agnostic for any type).
• Models can be ranked from low to high explainability, with Deep NN, Kernel-based (SVM), Clustering, Regression, and Classification.

Reliability and Resilience

• High data quality and security, robust algorithms, and sufficient validation are necessary for reliable and resilient systems.
• Human oversight offers an additional safety net.

Machine Learning and Simulation

• Machine learning can be used to assist simulation, and vice versa.
• Hybrid methods combine machine learning and simulation to improve performance.

Physics-Informed Machine Learning

• Physics-informed machine learning incorporates physical domain knowledge into the training process or the model, such as using known differential equations.
• This approach can improve model performance and accuracy.

Semi-Supervised Learning

• Semi-supervised learning increases available labeled data for training by assuming continuity, clusters, and/or manifold to improve model performance.

Active Learning and Incremental Learning

• Active learning and incremental learning are essential for dynamic systems where data becomes available gradually over time or exceeds memory limits.

Reinforcement Learning

• Reinforcement learning involves training a model based on rewarding desired behaviors and punishing undesired ones.
• Q-Learning is a type of reinforcement learning that tries to maximize the reward in each state depending on possible actions without a starting policy and without modeling a probability distribution.

Support Vector Machine (SVM)

• SVM involves classification by finding the decision boundary that maximally separates different classes.
• SVM is effective in high-dimensional spaces.

Neural Networks

• Neural networks can be used for classification, regression, and clustering tasks.
• Weights and biases are adjusted to improve the accuracy of neural networks using the gradient (from automatic differentiation or computational graph).

Evolutionary Algorithms

• Evolutionary algorithms mimic biological evolution and use fitness evaluation (objective function) for adaptation and convergence to find optimal solutions.

Firefly Algorithm and Particle Swarm Optimization

• Firefly algorithm emphasizes attraction (objective function) and randomness (diversity for exploration) to converge and escape local optima.
• Particle swarm optimization is a complex global optimization based on swarm dynamics (particles communicate about potential solutions and adjust search) based on the objective function.

Bayesian Optimization

• Bayesian optimization is an ML-based global optimization algorithm that uses Gaussian processes to approximate the objective function to search mainly in regions with the best expected result (good exploration-exploitation trade-off).
• Bayesian optimization is used for hyperparameter optimization to improve the accuracy of neural networks (layers, nodes, activation function).

Bias-Variance Tradeoff

• Bias occurs when a model is too simple and does not fit the data well (underfitting).
• High-variance occurs when a model is too complex, and small changes result in big changes in the solution (overfitting).

Transfer Learning

• Transfer learning involves learning new tasks based on previously learned tasks to improve the learning of the target predictive function using knowledge in the source domain and learning.
• Types of transfer learning include inductive, transductive, and unsupervised transfer learning.

Sim-to-Real Gap

• Training with simulated data is cheap, fast, scalable, and safe, but may not accurately model reality, leading to large control errors.

Ethics in AI

• AI should be fair, unbiased, transparent, and secure, while maximizing benefits and minimizing harm for humanity and the environment.

Safety and Security

• Vulnerability assessment and incident response must be taken into consideration to avoid susceptibility to attacks (poisoning, adversarial).
• Vulnerability can be minimized by appropriate model security, regular updates, and regulatory compliance, while a secure development environment needs to be ensured.