All Flashcards
What are the possible Objections or Arguments against intelligent machines?
- theology (have to own a soul?)
- “Heads in the sand”
- mathematics (uncomputability)
- consciousness
- various disabilites (“cannot do x”)
- lady lovelaces’s objection (“machines can only do what they are ordered to do”)
- continuity of nervous system
- informality of behavior
- extra-sensory perception
Difference between Extensional und Intensional Equality?
Intensional: A = A, d.h iff es handelt sich um dasselbe Objekt (Pruefen der Speicheradressen). Interne Gleichheit
Extensional: Wenn sie auf alle Eigenschaften auf die ich sie testen kann sich gleich verhalten (A = A’) Verschiedene Speicheradressen, aber gleiches Verhalten. Sicht und Test von Aussen auf Objekt
Worauf testet der Turing Test oder Imition Game? Extensional und Intensional Equality?
Extensional (=Wenn wir keinen Unterschied zu einer intelligenten Entität feststellen können, müssen wir sie wie eine intelligente Entität behandeln)
What says Church Turing Thesis?
Es ist nicht notwendig, verschiedene neue Maschinen für verschiedene Rechenprozesse zu entwickeln, denn Computer sind universell (abgesehen von der Geschwindigkeit). (Quantencomputer?)
Was ist das “Heads in the sand” Argument?
Man ignoriert die Moeglichkeit, das Maschinen so intelligent wie Menschen werden koennen. Axiom: Nur biologische Menschen koennen intelligent sein.
Consider the small grid-world shown in Example 1a. Define a set of possible observations O and a set of possible actions A in accordance to the definition given in the
lecture. For your definition, what are |A| and |O|?
Similarly, define a goal predicate γ1 that accepts a policy that (a) finds the way from
the start tile (S) to the goal tile (G) in the optimal amount of steps and (b) only steps on
tile-numbers which are strictly bigger than any of the tile-numbers visited before. Then
provide a solution path, i.e., πa, for which γ1(πa) = True. Which of the goal classes we
have covered in the lecture could γ1 fall under?
Now consider the grid world shown in Example 1b. For the given solution trajectory
πb (red line) give ⟨at⟩t∈Z and ⟨ot⟩t∈Z. What is the response from γ1(πb) in this case?
Recall that in the lecture we have covered two simple policy (search) approaches,
one of them being random search ρ(n). What is the chance of a policy π fulfilling the
goal predicate γ1 with ρ(4) for the grid world of Example 1a? What was the chance
of taking the trajectory shown in Example 1b? Finally, provide a tile-numbering and
goal-position in the Template 1c such that γ1(ρ(4)) is guaranteed to be True. You may
set one blocking tile ! that can not be traversed.
Describe briefly Goal Class 0, 1, 1.5, 2!
0: Keine Ziele (No Goals)
1: Ein Ziel/Akzeptanz/Kondition (Goal Predicate)
1.5: Multiple Ziele (Multiple Goal Predicates)
2: Mapping auf Rewards/ Evaluation mit Target Funktion (Goal Valuation)
Research and assign the terms phenotype, genotype, and fitness to this problem and give an example of each.
Give an example of each selection, mutation, and recombination functions. In this order, apply twice (for two steps). What is the effect on your population (size)?
How do Evolutionary Algorithms perform in comparison to the backtracking algorithm?
Worse than backtracking solution for small n on average, better for large n.
Explain what the behavior of the process Individual is. How would you program that behavior in a common programming language like Python or as behavior tree?
We would say that recombination within Individual is in-place. Why? How many children are generated from two individuals recombining?
In-place, because one of the parent individuals x or y is replaced by a new individual crossover(x, y). Therefore we change the process population in-place and without growing it in size even momentarily.
Mutation is not modeled in this process view of an individual’s life. Why might that be?
Process notation is based on interactions of processes. Since the mutation operator usually does not change/interact with other process objects, we would have to model the genome (-values) separately to show an effect happening.
Explanation: The process notation is based on interactions of processes, which means that each process object interacts with other process objects to produce an output. In the case of the evolutionary process, an individual’s genome is a process object that interacts with other process objects such as selection, crossover, and fitness evaluation to produce a new individual.
The mutation operator is usually applied to the genome of an individual to introduce variation and maintain diversity in the population. However, from the perspective of the process notation, mutation does not interact with other process objects directly. Instead, it modifies the genome of an individual, which then interacts with other process objects.
To model mutation in the process notation, we would have to treat the genome values as a separate process object and show how mutation changes them. This could be done by defining a mutation process that takes an individual’s genome as input and produces a modified genome as output. However, this would add complexity to the model and make it harder to analyze.
Therefore, in the process view of an individual’s life, mutation is often omitted to simplify the model and focus on the interactions between the main process objects. However, it’s important to note that mutation is a crucial mechanism in evolutionary algorithms and should not be ignored in the overall design and analysis of the algorithm.
2.3 Multi-Armed Bandit
Neural Networks 3.1 i)
Compute a forward pass for the input values x1 = x2 = 1 and the weights/biases shown in Fig. 1. What is the loss for this input?
Neural Networks 3.1 ii)
Neural Networks 3.1 iii)
Think about why this algorithm is called back-propagation? What makes this approach advantageous? (Hint: Think about the order of computation.)
Neural Networks 3.1 iv)
Bellman Equation 3.2
Multiple Choice 3.3
