Oraux
INFO-H-414 Swarm Intelligence
< RetourOral exam - 31 May 2022 - 3 Jun 2022
First part : my subject of choice (swarm robotics)Motivation behind automode vs evolutionary robotics (answer : reality gap)
Definition of swarm robotics
Definition Automatic design, advantages
Mostly the teacher just rebounded on what you said and “challenged” your thought process with counter example, starting from one very basic question, at least for me
Cost difference between swarm robotics vs normal robotics
I also talked about the “cache” system experiment, ants dropping leaves to the ground so that other ants can gather them. Labor division
Second part, after presentation on the project (swarm intelligence) :
Imagine you have a flight crew with pilots, stewards, etc, and planes, and itinerary, would you use PSO or ACO and how would you implement the algorithm to solve the issue
ACO vs PSO deterministic or stochastic (both are stochastic…)
What data structure would you use for PSO
The formula for PSO (each component)
Type of problem solved by ACO (shortest path problem)
Oral exam - 2 Jun 2022
my subject of choice was swarm robotis.first part was on swarm robotics:
q1 : what is swarm robotics?
--> See definitions in course slides
q2 : what is wilson experiment?
--> analyze the minor and major ants, talk about stimulus and treshold. Do not talk about the barnabeau experiments, i went from wilson directly to barnabeau and i think it negatively impacted my grade. Concentrate only on Wilson, it can be short explanation, if he wants more he will tell you.
q3: what is the difference between vanilla and chocolate?
-->the optimisation algo, in vanilla : F-Race, in chocolate : iterated-F-Race
q4: How did we know we had to change the optimisation algo in vanilla ?
--> In the 5 experiment with human, you have evo-stick (you don\'t care about this one for this question), vanilla, c-human, and u-human. What is important is to understand what is c-human, C stands for constraint. this is because c-humain have to use the same design as vanilla, the only difference is that there is no optimisation algo, it is the humans that decide how to optimize it. Since the only diff between c-human and vanilla is the algo and since c-human beat vanilla, the weak point of vanilla is it optimisation algo F-Race.
Second Part on swarm optimisation:
first i presented my project, it is the assistant that asked question. (project related question) since i had made my ground bot go in the center of the black surface, the assitant wanted to know and i told that i just made multiple test to find the correct distance manually. (i say this because marco\'s first question will be related to this).
q1: you say you implemented manualy the distance the robot should walk in the black area to go in the center of the area, how could you do this using an algorithm of swarm optimization, and what data structure would you have used?
--> algo : Particle swarm optimisation, data structure : vectors (for the particle, which each component of the vector being a parametre). How do you evaluate : you make many full simulation with different particle and take the one leading to the best results. (have to admit, strange question)
q2: in tsp you have some cities not connected what do you do ?
just make links between the cities with huge costs, the ants will thus make poor results with it and it won\'t be used in the end.
q3: in ACO you have alpha parametre and beta parametre, if alpha is equal to 0, so you don\'t have pheromon taken into account, how do you increase exploration?
--> decrease beta, and to have the most amount of exploration, put beta equal to 0 (uniform distribution in the end).
Oral Project + Course - 1 Jun 2022
Project:- What could be added to improve the algorithm ?
Swarm Robotics:
- Definition swarm robotics
- What is AuToMode and why was it introduced ?
- Link between scalability and local information
Swarm Optimisation:
- What is PSO ?
- What is ACO ?
- He gives an example and ask how can you resolve it to optimality (using PSO)
- What kind of data structure would be needed for PSO
- What can you do to have more Exploitation ? Exploration with ACO ?
Oral exam 2022 - 1 Jun 2022
(My in-depth topic choice was swarm robotics)Project:
What would be a good way to \"artificially\" increase the communication range of the robots ?
--> Build chains of robots as seen in the course
Swarm Optimization :
1) ACO and PSO: what kind of problems can you use them on ?
--> ACO: discrete optimization problems, PSO: continuous optimization problems (but there are variants that can deal with discrete problems as well)
2) ACO and PSO: deterministic or stochastic ?
--> Both stochastic, explain why.
3) What do the agents in ACO and PSO represent ?
--> The particles in PSO represent solutions themselves, but the ants in ACO only build the solutions
4)If you were solving TSP with ACO, and an ant had arrived in the last city and there was no road to first one to complete the cycle, what would you do?
-->Add an edge from last city to first one and make its cost super high so there is not much pheromone for this solution, and subsequent ants will not take the same path.
Swarm Robotics:
1) What is swarm robotics ? What are the advantages compared to classical robotics ?
--> See definitions in course slides
2) Explain some ways to do automatic design of robot swarm.
--> Speak about Evolutionary Robotics (ER) and Automode
3) What are the advantages of Automode over ER ?
ER is very sensitive to reality gap because NNs have high variance and thus have a tendecy to overfit during simulations.
31st may 2022 - 1 Jun 2022
I had chosen swarm robotics as my topic of preference.Reminder: don\'t forget to prepare the slides for you presentation!
First Part (project defence + swarm optimisation):
I was called by Marco Dorigo as well as 3 TAs. Having so many people listening to me was stressful so be prepared.
I was asked the following questions:
1. What is the role of the pheromones in ACO.
2. How can we give more importance to the pheromones?
3. You have a certain number of different chemicals at your disposition. How can you determine what amount to use of each of these chemicals in order to maximise a certain objective function?
The first two questions were straight to the point and theoretical (in the sense that the answer lies directly in the course). The last question was very tricky in my opinion. Basically he wanted me to tell that I shouldn\'t use ACO but rather PSO (because in ACO ants build the solution when in PSO the particles are the solution)
Second Part (swarm robotics):
I had to wait for more than 1h before being called by Mauro Birattari (as well as 2 TAs, again kind of stressful).
1. Define swarm robotics. What are the advantages of swarm robotics compared to a single robot? Give an example for the use of swarm robotics.
2. What is the difference between online and offline automatic design.
3. Can you talk about the techniques we saw during the lectures to implement automatic design?
1. The definition and advantages can be found in the course. An example would be a situation in which the loss of a robot would not compromise the mission.
2. and 3. The answers are directly in the course.
Mauro Birattari was very nice and reassuring.
Marco Dorigo was the complete opposite. He was very strict and unpleasant and at the end of the exam, when I asked him how I did, he laughed at my face before answering to my question.
Oral - 31 May 2022
The exam went kinda well they ask basic questions if you followed the class video! Just try to know what name of concept is linked to what. Because they don’t show you an equation and ask you to explain it they are asking you: “What is ant colony optimization?” “What is PSO?”.Sometimes they asks basic questions, so don’t hesitate and give basic answers.
That is what costed me some points I think.
For the project they observe your video and observe what are the key points where
your program could have done better.
Than they ask you questions about it and you have to tell what you could have done better!
Oral exam 2022 - 30 May 2022
My topic preference for this exam was swarm robotics.Part 1 (Swarm Optimization + project) :
Project :
- how could you have reduced interference ? (pattern formation)
Swarm-based Optimization :
- Are PSO and ACO stochastic or deterministic ? (both stochastic, he asked for the stochastic components in each)
- Exploitation/exploration in ACO (he seemed okay with me talking about evaporation)
- What are the agents in ACO vs PSO (ants : build a path and solutions but are not solutions, particules are solutions)
Marco Dorigo is very picky but does somewhat help.
Part 2 (Swarm Robotics) :
- Definition of swarm robotics and what are the properties of a swarm, + and - against classical robotics, challenges
- Definition and explanation of stigmergy
- Division of labor in insect societies, address plasticity
- Automatic design approaches (ER & Automode), what are the pros and cons of each other (talk about the overfitting ER does on the simulation, making it harder for ER to cross the reality gap, explain that automode injects biases that help maintain a good performance with real robots)
- Applications of swarm robotics (asked by an assistant, I talked about mapping and a bit of search and rescue but I feel like Mauro was already satisfied)
Mauro is super nice and helps/clarifies if needed.
INFO-H-414 Swarm Intelligence - 16 Jan 2016
Part 1: Oral-Data clustering: principles and equations
-Differences between PSO and ACO. PSO is improving solutions, ACO is building solutions.
-Equations of PSO and ACO.
-PSO: Identify the exploration and exploitation parts. For instance, the random matrix U is pushing toward exploration.
-ACS: equations and explain everything (like why is alpha = 1?, what is that q_0?, etc.)
Part 2: Project
-Presentation
-Questions on the project (is it applicable IRL? what to improve, does it use swarm robotics principles like desynchronization of agents and stochastic behaviour?)
-Wilson model
-DoL: what is it? => Engaging a task based on the stimulus, that's it.
-Threshold model: give it, explain it, equations, graphs, principles.
All in all, Mauro Birattari is nice and helps a lot, but he is ruthless. Marco Dorigo is passive, silent, does not help a bit, but the teaching assistant does (well, depending on the TA, I guess).
Il n'y a pas de publications plus anciennes.