p(B|A) = p(B and A) / p(A)

= p(A|B) * p(B) / p(A)

• Describe which variables influence which otehr variables
• No connection between two variables implies conditional independence

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

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

• 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

• Agent lives in an environment, it must choose actions within that world and periodically it gets either positive or negative reinforcement

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):

V_i = V_i + a [r + dV_j - V_i]

• Different from Value per state V_state
• We keep track of a Q value for each possible state-action pair
• This is also called the model-free learning

_Qi,k = Q_i,k + a [r + d.max(Q_j,x - V_i,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

• 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

• Biological concept
• Way organisms become better suited to their environment over time
• "Survival of the fittest"

• 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

1. Choose a random solution
2. Generate a mutated variant (e.g. flip a random symbol to some other value
3. If the variant is better, it replaces it and then we repeat step 2
4. If there are no improvements after N tries, go back to step 1 (but always remember best solution so far)

1. Variation: Individuals not all the same; some random variation present
2. Selection: not all individuals survive to reproduce, and some individuals reproduce more than others
3. Heredity: individuals tend to be like their parents

• Population of solutions/individuals
• Mapping from genotype to phenotype
• Fitness function
• Selection procedure
• Crossover
• Mutation

• 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

Idea is that advantageous mutations can be shared around population 

• Single-point
• Multi-point
• Uniform

• 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

• 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)