AI & Game Programming QUIZ for Mid-term

Quiz A

1. What is not covered in kinematic movement algorithms?

1. Position
2. Orientation
3. Velocity
4. Rotation
5. Accelerations

2. What is the difference between the seek and arrive algorithms?

1. Movement of the target
2. Direction of the target
3. Presence of the target

3. How to prevent overshooting in the arrive algorithm of steering behaviors?

1. Use variable matching with target velocity
2. Use two radii to slow down and stop

4. In the obstacle and wall avoidance of steering behaviors, casting the charcater's rays at a large angle can avoid corner traps.

1. O
2. X

5. A* with zero heuristic is the pathfinding version of Dijkstra

1. O
2. X

6. Dijkstra will never find a better route to a closed node

1. O
2. X

7. Dijkstra searches the entire graph indiscriminately for the shortest possible route.

1. O
2. X

8. Choose difference between Dijkstra and A*

1. using heuristic value
2. updating node (closed, open) list
3. comparing value when updating
4. terminating way

9. When heuristic is underestimated, then

1. A* will prefer to examine nodes closer to the start node
2. The estimated-total-cost will be biased toward the heuristic value

10. When heuristic is overestimated, then

1. Algorithm will tend to favor nodes that have less distance to go
2. It will be the exact same path that the Dijkstra algorithm would generate

11. Which heuristic is better when the map is indoor?

1. Euclidean Distance
2. Cluster
3. Dijkstra (none)

QUIZ B

• For classification problem (number of class: K), the softmax output y is probability that input x belongs to each class

• MSE loss function이 Cross Entropy보다 더 수렴을 잘 한다

• Deep learning process with multiple layers, final output contains the most relevant features and noise is filtered out

• When solving a task with sufficient data, a deeper network requires more neurons than a shallow network

• When there is no activation function, the MLP is just learning linear transformation of input data

• What happen when we initialize the weight with high values?

• For better convergence, we need to do finer search of learning rate

• What are the methods to achieve better convergence without falling into local minima?

• A single perceptron can solve nonlinear problems like the XOR problem (O/ X)
• For X with d=1 and output y<0, choose the update formula for w

1. w=w+x 2. w=w-x 3. w=w*x 4. w=w/x 5. w=w(1-x)

• For linearly separable binary data, the perceptron learning algorithm can obtain t
  he weight vector w with a finite number of updates (O/ X)

• Most real-world data is not linearly separable and contains many outliers, necessitating the use of MLPs to deal with such data and to find weights that minimize the error (O/ X)

• A perceptron composed of an input layer and an output layer can be considered as a 2-layer MLP (O/ X)

• When finding optimal weights in MLP, using a random guess is an effective method initially (O/ X)

• The sigmoid function is more suitable as an activation function than the step function because it can have a wider range of output values (O/ X)

• MLPs cannot solve multi-class classification problems (O/ X)

• Choose back propagation method that divides the data into small parts and uses some of it for training. (SGD/ BGD/ MGD)

답은 알아서 ㅎㅎ