Term
What is the bayesian formula? |
|
Definition
|
|
Term
|
Definition
p(A|B) = p(A and B) / p(B) |
|
|
Term
What does p(A and B) mean? |
|
Definition
p(A and B) = p(A|B) * p(B) |
|
|
Term
|
Definition
p(B|A) = p(B and A) / p(A)
= p(A|B) * p(B) / p(A) |
|
|
Term
Show a joint distribution of 3 variables |
|
Definition
|
|
Term
How can you reduce (the variables of) a joint distribution table? |
|
Definition
If a variable is independent there is no need to compare it to each of the other variables as if A is independent of B, then p(A|B) is just equal to p(A) |
|
|
Term
What is a bayesian network used for? |
|
Definition
- Describe which variables influence which otehr variables
- No connection between two variables implies conditional independence
|
|
|
Term
|
Definition
|
|
Term
|
Definition
Decision support tool that uses tree like graph or model of decisions and their possible consequences, including chance event outcomes, resource costs, utility, etc. |
|
|
Term
|
Definition
The measure of the amount of disorder or surprise in a system
High entropy means we have no idea what is going to happen.
Low entropy means we've pinned thiings down to some extent; we've got some information about what is likely to happen |
|
|
Term
What is the entropy formula? |
|
Definition
H(X) = Ex(I(x)) = -sum(p(x)log2(p(x)) for all x |
|
|
Term
How to build a decision tree? |
|
Definition
Split the tree on the attribute with the highest information gain. Then recurse |
|
|
Term
What is the k-neighbour algorithm? |
|
Definition
http://www.youtube.com/watch?v=4ObVzTuFivY
- It takes a specific point, and classifies it according to the majority vote of the k nearest points
|
|
|
Term
What is supervised learning? |
|
Definition
- The learner must learn to classify cases but a labelled training set spells out what the right answer should be in each case.
- K-nearest neighbour, decision trees, neural nets are all examples of supervised learning
|
|
|
Term
What is reinforcement learning? |
|
Definition
- Agent lives in an environment, it must choose actions within that world and periodically it gets either positive or negative reinforcement
|
|
|
Term
What is the temporal difference learning equation? |
|
Definition
Vi = Vi + a [r + Vj - Vi]
Where
- Vi = new opinion
- Vj = old opinion
- a = learning rate
- r = actual reward
|
|
|
Term
What is discounting future rewards and how can it be used? |
|
Definition
Rewards that can be obtained now are usually better than the one we obtain later
We need to discount the future rewards implicit in Vj with a new d term (e.g. d= 0.9):
Vi = Vi + a [r + dVj - Vi] |
|
|
Term
|
Definition
- Different from Value per state Vstate
- We keep track of a Q value for each possible state-action pair
- This is also called the model-free learning
|
|
|
Term
What is the update formula for the Q-learning? |
|
Definition
Qi,k = Qi,k + a [r + d.max(Qj,x - Vi,k]
Where
- We're in state i, we choose action k that takes us to state j and gives us reward r
- Learning rate a = 0.1, discount factor d = 0.9
- Expected value of getting to state j is the maximum Q value we could get for any action x done at j
|
|
|
Term
What are the characteristics of local search? |
|
Definition
- Local search methods are highly general
- Means starting somewhere in a space of posibilities and iteratively trying the neighbours of our current location to see if they are better - thus myopic
- Used when we're ignorant of the global structure of our possiblity space
- Computationally inefficient
- Mirrors natural adaptation
|
|
|
Term
|
Definition
- Biological concept
- Way organisms become better suited to their environment over time
- "Survival of the fittest"
|
|
|
Term
|
Definition
- If value for one genetic value depends on the values of the other variables
- AKA epistasis (biology) interaction (Statistics), frustration of variables (engineering)
- As local correlation goes down the ruggedness of the landscape goes up and becomes more difficult to search
- Rugged landscapes are high in epistasis, whereby the fitness contribution of one parameter is modulated by others
|
|
|
Term
What is the random-restart hill climber? |
|
Definition
- Choose a random solution
- Generate a mutated variant (e.g. flip a random symbol to some other value
- If the variant is better, it replaces it and then we repeat step 2
- If there are no improvements after N tries, go back to step 1 (but always remember best solution so far)
|
|
|
Term
What are the features of genetic algorithm? |
|
Definition
- Variation: Individuals not all the same; some random variation present
- Selection: not all individuals survive to reproduce, and some individuals reproduce more than others
- Heredity: individuals tend to be like their parents
|
|
|
Term
What are the components of a Genetics algorithm? |
|
Definition
- Population of solutions/individuals
- Mapping from genotype to phenotype
- Fitness function
- Selection procedure
- Crossover
- Mutation
|
|
|
Term
What are the principles of Selection in a GA? |
|
Definition
- Individuals with higher fitness scores are more likely to reproduce
- Can be achieved through fitness-proportionate "roulette-wheel" selection
- Rank-based selection and touranment selection also used
|
|
|
Term
What are the principles of Crossover in a GA? |
|
Definition
Idea is that advantageous mutations can be shared around population
- Single-point
- Multi-point
- Uniform
|
|
|
Term
What are the principles of Mutation in a GA? |
|
Definition
- the source of variation
- Without mutation, crossover couls shuffle the initial set of genes around, but there would be no evolutionary novelity
- In a bit-string, mutation is implemented as a probability that any one bit will be flipped during reproduction
- Things get trickier with real-valued genotypes, and a poorly chosen mutation operator may introduce biases
|
|
|
Term
What are the principles of the law of uphill analysis and downhill invention? |
|
Definition
- Analysis - figuring out the vehicle's circuitery from outside is much tougher than
- invention - playing around with different vehicle designs to see what they doo
- A psychological consequence:
- We tend to overestimate the complexity of the mechanisms behind cognitive systems (e.g. we propose a complex modular architecture)
|
|
|