You are on page 1of 8

Sistemas Autónomos/Autonomous Systems

1º Semestre 2016/17

MEEC, MEAero, MEIC


2º Teste
15 de Dezembro de 2016

NUMBER _______________
NAME_______________________________________________________________

1. Logic has been used in many Artificial Intelligence and Robotics systems for
knowledge representation and inference. Which of the following combined statements
is true?
 a) Logic systems are composed of a syntactic system and a semantic system.
The syntactic system is decomposed into the alphabet + formation rules
(constituting the language) and the inference rules.
 b) Logic systems are based on a syntactic system, which is decomposed into
the alphabet + formation rules (constituting the language) and the
inference rules.
 c) Logic systems are based on a semantic system, which is decomposed into
the alphabet + formation rules (constituting the language) and the
inference rules.
 d) Logic systems are composed of a syntactic system and a semantic system.
The semantic system is decomposed into the alphabet + formation rules
(constituting the language) and the inference rules.

2. The use of predicate logic for task planning


 a) Is not possible.

 b) Is possible with extensions such as those introduced by Situation Calculus


and STRIPs.
 c) Is possible by adding a time variable to handle situation changes.

 d) Is possible by adding quantifiers, variables, objects, functions and relations


to handle state changes.

Each question have an associated grade of 1,0 point for correct answers, -0,1 points for incorrect answers and
0 points for no answer.
Mark an “X” on the  corresponding to the correct answer. In case you change your mind about the correct
answer later, cross the former “X”, and mark another “X” corresponding to your new decision.
1/8
3. When planning in a dynamic and uncertain environment, which of the following
assumptions is made in the original formulation of the basic logic-based planning
problem?
 a) Agent does not know everything that is relevant for the planning problem.
 b) Agent knows how its available actions can change the world state from one
state to another.
 c) The planning agent is not in control of the world – state changes are the
result of its deliberate actions or of the actions of other agents.
 d) The agent’s preferred world states (or the goal states) may change during a
planning episode.

4. One of the extensions of the basic logic-based planning problem is continual


planning, defined in this course as:
 a) Handling uncertainty by enumerating the possible states that may arise after
the execution of an action and providing alternative courses of action for
each of them.
 b) Monitoring progress during plan execution and, when deviations from the
predicted nominal conditions occur, halting the plan execution and creating
a revised plan.
 c) Revising the plan as an ongoing process rather than one triggered by
failures of the nominal plan. Planning is not made in too much detail into
the future, and it is interleaved with execution.
 d) None of the above.

5. In the basic planning problem formulation (LaValle, 1996), a plan is represented by:
 a) An undirected graph, which vertices represent states and an edge connecting
state x and state x’ exists in the graph if there exists an action a such that
x’ = f(x,a), where f is the state transition function.
 b) A Petri net graph, which places represent states and an arc from state x to
state x’ exists in the graph if there exists an action a such that x’ = f(x,a),
where f is the state transition function.
 c) A directed graph, which vertices represent actions and an edge from action
a to action a’ exists in the graph if there exists a state x such that a’ = f(x,a),
where f is the action transition function.
 d) A directed graph, which vertices represent states and an edge from state x to
state x’ exists in the graph if there exists an action a such that x’ = f(x,a),
where f is the state transition function.

2/8
6. Following the rationale of the basic planning problem formulation (LaValle, 1996),
if each state is represented by the combination of Boolean values of 5 predicates, 3
of them with 2 arguments, and the other 2 predicates with 1 argument each, and also
assuming that each argument can be instantiated by 4 possible values, the dimension
of the state space is
2
 a) 23∗4 ∗ 22∗4
2
 b) 25∗4
2
 c) 22∗4 ∗ 23∗4
 d) 3 ∗ 42 + 2 ∗ 4

7. The Stanford Research Institute Problem Solver (STRIPS) models the effect of
actions (operators under STRIPS terminology) over the world as:
 a) A probabilistic action model p(x’|x,u), where x is the current state
(situation), x’ the resulting state (situation) and u the applied action
(operator).
 b) A deterministic action model x’ = f(x,a), where x is the current state
(situation), x’ the resulting state (situation) and u the applied action
(operator), and f is a continuous transition function that maps <state,
action> pairs into <states>.
 c) A deterministic action model x’ = f(x,a), where x is the current state
(situation), x’ is the resulting state (situation), u is the applied action
(operator), and f is a table that, for a given <state, action> pair, removes
logical clauses from, and adds logical clauses to, the current state to build
the new state – states being represented as a set of clauses.
 d) STRIPS has no model of action (operators) effects over the world.

8. A Markov Decision Process (MDP) models the effect of actions over the world as:
 a) A probabilistic action model p(x’|x,u), where x is the current state, x’ the
resulting state and u the applied action.
 b) A deterministic action model x’ = f(x,a), where x is the current state, x’ the
resulting state and u the applied action, and f is a continuous transition
function that maps <state, action> pairs into <states>.
 c) A deterministic action model x’ = f(x,a), where x is the current state, x’ is
the resulting state, u is the applied action, and f is a table that, for a given
<state, action> pair, removes logical clauses from, and adds logical clauses
to, the current state to build the new state – states being represented as a set
of clauses.
 d) An MDP has no model of action effects over the world.

3/8
9. Which of the following Petri nets adequately represents a task plan executed by two
robots which at some point needs to synchronize the execution of the two plans (i.e.,
robot 2 must wait for robot 1 to execute some part of its plan):
 a)
Robot 1 Robot 2
task plan task plan

 b)
Robot 1 Robot 2
task plan task plan

 c)
Robot 1 Robot 2
task plan task plan

 d)
Robot 1 Robot 2
task plan task plan

4/8
10. A Markov Decision Process (MDP) is formulated as:
 a) A finite (or countably infinite) set of states, a finite set of actions, a
deterministic transition function mapping <state, action> pairs into states,
and a reward function mapping <state, action> pairs into a real number.
 b) A finite (or countably infinite) set of states and a probabilistic transition
function between states.
 c) A finite (or countably infinite) set of states, a finite set of actions, a reward
function mapping <state, action> pairs into a real number and a performance
function, usually the expected discounted sum of rewards obtained by the
application of actions, over a finite or infinite time window, starting from
the current state.
 d) A finite (or countably infinite) set of states, a finite set of actions, a
probabilistic transition function mapping <state, action> pairs into states
with a given probability, a reward function mapping <state, action> pairs
into a real number and a performance function, usually the expected
discounted sum of rewards obtained by the application of actions, over a
finite or infinite time window, starting from the current state.

11. Consider the following marked Petri net:

t1
p3
2
t2
2
p1
p4

t3
p2

When firing the transition sequence t3 t2, which of the following marking sequences
describes the corresponding state trajectory of the Petri net?
 a) (1 1 0 1)  (1 1 0 2)  (2 1 1 0)
 b) (2 1 0 1)  (1 1 0 2)  (2 1 1 0)
 c) (2 1 0 1)  (1 0 0 2)  (0 0 1 0)
 d) the transition sequence is not feasible

5/8
12. A possible formulation of the event concept, used in class under the framework of a
robot task model, and its association to a Petri net (PN) representation of a task plan,
is:

 a) An event is controllable if it represents uncertain action effects, and is


associated to PN transition(s).
b) An event is the result of a change of Boolean value of a predicate (after
its arguments are instantiated), and can be associated to a PN place;

equivalently, the same predicate can be associated to an output transition
of the referred place.
c) An event is the result of a change of Boolean value of a predicate (after
its arguments are instantiated), and can be associated to a PN transition;

equivalently, the same predicate can be associated to an input place of the
referred transition.
 d) None of the above.

13. A Generalized Stochastic Petri net (GSPN) [choose the most general description]:

 a) Has timed stochastic transitions, with an associated function that maps the
transition in each reachable marking to a random variable representing the
firing time of the transition, and immediate transitions. When a subset of
the immediate transitions is enabled in a given marking, a probability mass
function is associated to the subset by random switches.

 b) Has immediate transitions, with an associated function that maps the


transition in each reachable marking to a random variable representing the
firing time of the transition, and stochastic transitions. When a subset of
the stochastic transitions is enabled in a given marking, a probability mass
function is associated to the subset by random switches.

 c) Has timed stochastic transitions, with an associated function that maps the
transition in each reachable marking to a random variable representing the
firing time of the transition.

 d) Has timed stochastic transitions, with an associated function that maps the
transition in each reachable marking to a random variable (exponentially
distributed) representing the firing time of the transition, and immediate
transitions. When a subset of the immediate transitions is enabled in a
given marking, a probability mass function is associated to the subset by
random switches.

6/8
14. Which of the following is an example of a quantitative stochastic design problem for
a robot task plan modelled by a Petri net:
 a) Ensure that the plan will not lead to a deadlock, when the algorithm
 implementing one of the plan primitive actions is replaced by a new one

 b) Ensure that all resources necessary to execute the plan will not be over-
requested, with probability 1.
 c) Design a supervisor that ensures the plan will not take longer than 60 minute.
 d) Ensure that the plan will achieve a probability of success above some given
threshold, when the probability of failure of one of its primitive actions
increases 10%.

15. Which of the following statements is true regarding Markov Decision Processes
(MDPs):
 a) Solving an MDP by Dynamic Programming does not require the a priori
knowledge of transition probabilities and rewards.
 b) Solving an MDP by Dynamic Programming has linear computational
complexity with respect to the number of states.
 c) Solving an MDP by Dynamic Programming requires the a priori
knowledge of transition probabilities and rewards.
 d) Solving an MDP by Dynamic Programming requires only the a priori
knowledge of transition probabilities, but not of rewards.

16. In a Markov Decision Process, the relation between the Q value of a <state, action>
pair Q(s,a) (used in Q-learning) and the state value V(s) is:
 a) 𝑉(𝑠) = max 𝑄(𝑠, 𝑎)
𝑎

 b) 𝑄(𝑠, 𝑎) = max 𝑉(𝑠)


𝑠

 c) 𝑉(𝑠) = min 𝑄(𝑠, 𝑎)


𝑎

 d) 𝑄(𝑠, 𝑎) = min 𝑉(𝑠)


𝑠

17. In the Q-learning algorithm, the trade-off between exploration and exploitation when
selecting the action at each step, aims to ensure that:

 a) Each pair <state, action> is visited a finite number of times, so that the
estimate of Q converges to the real value of Q.

 b) Each pair <state, action> is visited an infinite number of times, so that the
best action is always selected.

 c) Each pair <state, action> is visited an infinite number of times, so that the
estimate of Q converges to the real value of Q.

 d) Each pair <state, action> is visited an infinite number of times, so that a


global maximum of Q is reached.

7/8
18. The epsilon-greedy method for action selection in random search algorithms
(including Q-learning) selects actions at each step of the algorithm by:
 a) Picking one of the actions with equal epsilon probability.

 b) Picking the best action so far with probability 1-epsilon, and picking with
probability epsilon, uniformly distributed over the remaining actions, one
of those actions.
 c) Picking one of the actions with an exponential probability mass function
with rate epsilon, defined over the set of actions.
 d) Picking one of the actions with a Gaussian probability mass function with
mean value equal to epsilon, defined over the set of actions.

19. Solving the sensor fusion problem to estimate the position of a static object requires
that:
 a) A probabilistic model of the object motion is defined in the world frame.

 b) Gaussian probability density functions representing the uncertainty about


the object position refer to some frame.

 c) All probability density functions representing the uncertainty about the


object position must be expressed on the each sensor’s body frame.
 d) All probability density functions representing the uncertainty about the
object position must be expressed on the same (world) frame.

20. The Linear Opinion Pool (LOP) method for sensor fusion combines sensor
measurements by calculating a probability mass function (pmf) p:

 a) 𝒑 = ∑𝑖 𝑤𝑖 𝒑𝑖 , where pi is the pmf of sensor i, and the summation weight wi


is proportional to the determinant of the covariance matrix of sensor i
measurement model.

 b) 𝒑 = ∑𝑖 𝑤𝑖 𝒑𝑖 , where pi is the pmf of sensor i, and the summation weight wi


is inversely proportional to the determinant of the covariance matrix of
sensor i measurement model.

 c) 𝒑 = ∏𝑖 𝑤𝑖 𝒑𝑖 , where pi is the pmf of sensor i, and the summation weight wi


is proportional to the determinant of the covariance matrix of sensor i
measurement model.

 d) 𝒑 = ∏𝑖 𝑤𝑖 𝒑𝑖 , where pi is the pmf of sensor i, and the summation weight wi


is inversely proportional to the determinant of the covariance matrix of
sensor i measurement model.

8/8

You might also like