Övrigt Massa Bra Flashcards
Overshooting
When the learning rate of the model is too high so that the optimal point is missed and prevents the model from converging to minimal loss and makes it instable
Overfitting
Model is overtrained on the trainingdata, it learns noise and other factors. It can no longer generalize and will result in a decrease in loss of training data but an increase in loss of evaluation data
Underfitting
The model is not trained on enough data and will result in a fail of training set
Sigmoid Function vs Heavside
A sigmoid activation function will result in a output between 0 - 1. It is best used for say classification where the output then resembles the certainty/probability.
Heavside on the other hand results in either 0,1, good for binary classifications or boolean values
Temporal Difference
Temporal Difference Error is the difference between the observed Value of a state and the expected. value
Difference Between DSS and ES and KBS
DSS contains a model base of mathematical models
ES and KBS contains knowledge bases and inference engines
ES and KBS are designed to emulate a domain expert
DSS can work in any domain
KBS is modular and its knowledge base can be exchanged, in difference to ES
Differences Database and KB
A database contains facts and files of information, internal and external data. But the representation and use is limited in comparison to KBs.
A knowledge base also stores facts but also contains rules for applying the knowledge and relations about the fact.
Knowledge stored in the KB can be represented in many different way
- As production rules, heuristics or rules of thumb, logic, pseudo code etc.
A knowledge base stores knowledge specifically represented for use in intelligent agents and systems.
Pruning of NN?
Pruning is a method where the network drops 50 % of the neurons in the hidden layers to avoid overfitting
Performance Measure of Agents
How a Agent is measured on their performance, eg. A goal based agent is easy to evaluate based on if the goal is fulfilled or not
What are requirements of intelligent Agents?
Autonomous, problem solving, reactive to their environment, ability to adapt etc. Depending on.
What are factors of problem domains
Time Complexity, Space complexity, Branching factor, problem instance and problem space
Machine Learning?
Machine Learning is the training of machines to find and recognize patterns in data
Deep Q Networks?
Deep Q Networks utilizes Deep neural networks to approximate Q-value function
Temporal Difference learning Algorithm
A learning algorithm that is used to find the best paths to the goal.
Learning rates
The size of adjustments of weights
