lecture 4

Question 1

Dennett’s cognitive wheel

Answer 1

The fact that you can invent something that works does not guarantuee that you have learned something about nature.

Question 2

Intelligence

Answer 2

Ability to deal with difficult and novel problems

o Adaptivity
o Very close to creativity

Question 3

Time scales of adaptive behavior

Answer 3

1. Extremely slow (evolutionary)
   e.g., cuckoo
2. Slow (operant conditioning)
   e.g., skinner box
3. Fast (problem solving)
   intelligence

Question 4

Think aloud protocols

Answer 4

Write down what people think during chess and replicate these processes

Question 5

Drosophilia idea

Answer 5

o By building AI for chess, we hopefully learn how chessplayers think

o This failed.

Question 6

why is chess difficult

Answer 6

1. Explosion of possibilities (novel positions)
   a. Combinatorial explosion
2. What is a good position?

Question 7

simple problems

Answer 7

Problems for which the time to solve the problem does not increase fast when the problem becomes bigger

 Polynomial time

Question 8

hard problems

Answer 8

The solution times increases very fast
o Takes infinite amount of time
o Same with chess

 Bc it checks all the positibilities
 Non-polynomial time
 If you solve one, you solve everything

Question 9

classic AI solutions to solving problems

Answer 9

1. Tree algorithms
2. Evaluation function
3. Build in a huge opening book
4. End game table bases

Question 10

Alpha-beta pruning

Answer 10

Stops evaluating a move when at least one possibility has been found that proves the move to be worse than a previously examined move

so they don’t follow the whole tree

this way they can go much deeper

Question 11

deep blue

Answer 11

 Hardware + 4 tricks (previously mentioned)
 Kasparov lost against Deep Blue

is a cognitive wheel because this is not how humans solve problems/learn

Question 12

Learning in AI

Answer 12

1. Deep learning
   a. Supervised and unsupervised learning
   b. Neural networks
2. Reinforcement learning
   a. Learn interations that are rewarded

Question 13

Q learning

Answer 13

 Fill out the q-table by randomly walking and sometimes it gets a reward. This creates the fastest pathway to the goal.
 There is also exporation vs exploitation in these tables
 Having sub goals also helps

Question 14

Deep reinforcement learning

Answer 14

 Q-table is replaced by deep learning NN that predicts next moves and learns from rewards

Question 15

Monte Carlo Tree search

Answer 15

 Start with random move and play out the game entirely with random moves
o Rollout

 Do this multiple times and use average result to add values to moves A and B

 When MCTS combined with deep reinforcement learning, it gives very powerful learning

> no evaluation of positions, just win/draw/loss at the end of a rollout

Question 16

minimax tree

Answer 16

search all paths until a certain depth and then use the evaluation function

Question 17

alphazero chess

Answer 17

o Self-play
 Played against itself all the time
o Reinforcement learning
o Deep neural network
o MCTS

Question 18

Bongard problems

Answer 18

humans generally outperform AI systems on these types of AI fluid intelligence tests

Question 19

brute force approach to NP problems

Answer 19

successful but not informative about human intelligence