====== Symbolické strojové učení ======

  * Stránky předmětu: [[https://cw.fel.cvut.cz/wiki/courses/B4M36SMU/start|Symbolické strojové učení]]
  * Přednášející: Jiří Kléma, Filip Železný
  * Cvičící: Jáchym Barvínek, Ondřej Hubáček, Petr Ryšavý, Martin Svatoš

===== Cvičení =====



===== Zkouška =====

06.06.2018
  - (**5 pnts**) Difference between PAC-learning agent and mistake-bound agent. 
      * (2 pnts) What does it mean when an agent in both frameworks learns? 
      * (3 pnts) What does it mean when it learns efficiently? Online?
  - (**10 pnts**) Space-version agent. There are given two agent with different hypotheses spaces. First is all possible 3-conjunctions (non-negative) of n variables. Second is all n-conjunctions of positive and negative literals. 
       * (3 pnts) For each agent: does it learn online? 
       * (3 pnts) For each agent: does it learn efficiently? 
       * (4 pnts) For the first agent: given the first negative observation (0,1,1,1,...,1), what will be the agent's decision on the next observation (0,1,0,1,...)?
  - (**15 pnts**) Relative Least General Generalization (rlgg). Given background knowledge B = {half(4,2), half(2,1), int(2), int(1)}. What will be the rlgg of o1 = even(4) and o2 = even(2) relative to the background? 
      * (10 pnts) Apply algorithm, draw tables, theta functions. 
      * (5 pnts) Make a reduction step relative to B. Why is it needed?
  - (**10 pnts**) Bayesian networks. 
     * (2 pnts) Find optimal, efficient, complete network (something like Season -> Temperature -> (two children: -> Ice Cream Sales, -> Heart Attack Rate)). 
     * (2 pnts) Then compute CPT (conditional probability tables). 
     * (3 pnts) Compute Pr(Spring|Good Ice Cream Sales, No Heart Attack)
     * (3 pnts) Compute Pr(Heart Attack|Winter, Bad Sales).
  - (**5 pnts**) Q-learning. Given 5 small questions, response True/False and provide your reasoning. 
     * (1 pnt) Can Q-learning be extended to infinite states or action space? How would it handle this?
     * (1 pnt) Does Q-learning use on-policy update? What is the difference from off-policy update?
     * (1 pnt) Does Q-learning always converge? If so, is it conditioned by anything? By what?
     * (1 pnt) Is Q-learning just an instance of temporal difference learning? If not, what is different?
     * (1 pnt) What is the difference between Q-learning and direct utility estimation or adaptive dynamic programming? What is better? 
  - (**5 pnts**) Q-learning representation. 
      * There is a robot moving in a swimming pool, which can move in either of 3 dimensions and it has exactly one propeller for each dimension. It can also move with two different speeds. There is a treasure at a specific place and a specific depth. There are mines at some places as well. If the robot hits a mine or the wall, it restarts at a random position.
      * (3 pnts) Describe states, actions, rewards of a specific game. You may provide two different representations.
      * (2 pnts) Describe Q-learning representation, the update rule, gamma, alpha value. How are Q values defined?

~~DISCUSSION~~