COM00177M
MEng and MMath Degree Examinations 2023–24
DEPARTMENT OF COMPUTER SCIENCE
Evolutionary & Adaptive Computing
Open Individual Assessment
Issued: 12 noon on 22 February 2024
Submission due: 12 noon on 23 May 2024
Feedback and marks due: 27 June 2024
All students should submit their answers through the electronic submission system:
http://www.cs.york.ac.uk/student/assessment/submit/ by 12 noon on 23 May 2024. An assessment that has been submitted after this deadline will be marked initially as if it had been handed in on time, but the Board of Examiners will normally apply a lateness penalty.
Your attention is drawn to the section about Academic Misconduct in your Departmental Handbook: https://www.cs.york.ac.uk/student/handbook/.
Any queries on this assessment should be addressed by email to Simon O’Keefe and Dimitar Kazakov at [email protected] or [email protected]. Answers that apply to all students will be posted on the VLE.
Rubric
Your submission should be a single zip file named after your exam number, Yxxxxxxx.zip, which contains:
• For Part 1 of this assessment
– a single Jupyter or Colab notebook EVAC1.ipynb combining code and explanations
– a PDF EVAC1.pdf of the state of the same notebook after all of its code has been executed.
• For Part 2 of this assessment
– a pair of NetLogo files EVAC2-train.nlogo and EVAC2-test.nlogo, possibly accompanied by additional image and data files. The first of these files should allow to run the evolutionary algorithm (with a relevant
description in its Info section). The second file should contain an already evolved behaviour for every dog, and allow for the performance of the
evolved herding dogs to be tested. It should also contain an Info section
with a report on the results and all remaining comments.
If there are discrepancies between the contents and output of the Jupyter/Colab notebook and what is shown in the PDF, the former will be used for marking.
Do not use archive formats other than zip. Your submission MUST be a single zip archive. Your code should assume any auxiliary files that you may generate are in the same folder as the notebook from which they are accessed.
1 Evolution of a function (this part is worth 50 marks)
1.1 Scenario
The scenario for this part of the assessment is the search for routes through a set of cities, subject to multiple constraints. A sales repreresentative is visiting cities to make sales, and has a good estimate for the sales to be made in each city. Visiting every city will maximise sales, but also maximise the distance travelled. The aim is to minimise the distance travelled, which may be to a subset of cities, but still make a lot of sales. The dataset contains a matrix with the distance between each city, and a list of the value of sales in each city.
1.2 Task
This part of the assessment requires you to use an evolutionary algorithm to evolve routes that start at a particular city, visit a number of other cities, and return to the start city. Apart from the start city, each city may only be visited once. The start city may be any city–that is, any route around any subset of cities is valid. The objectives are to:
1. Minimise the distance travelled
2. Maximise the value of sales made (counting the start city only once)
You must follow these rules of implementation:
(a) You MUST use Python and the DEAP library to implement an evolutionary algorithm.
(b) You MUST submit a Colab or Jupyter notebook containing runnable code for your evolved solution.
(c) You MUST evaluate thoroughly the performance of the evolved solution as appropriate to the problem.
1.3 Data
The data for this assessment may be found on the Assessment page of the module VLE site.
1.4 To do
You should report on this part of the assignment by adding comments to the Colab/Jupyter notebook you have created.
The marking criteria assume there is working code. Partial marks may be allocated for design alone. No results will be accepted if the corresponding parts of the code involved in their generation cannot be executed.
1. Description of the algorithm implemented [15 marks]
Give a full description of the chosen representation for individuals, and of the algorithm that you finally have chosen. You should give a clear justification for all the design choices made.
2. Quality of code [5 marks]
You must provide runnable code and a solution for the evolutionary algorithm using Python and DEAP. Your code should be well commented, and demonstrate the competent use of features of DEAP.
3. Investigation of parameters and representation [10 marks]
You should make a systematic investigation of the effect of parameters and of the choice of representation in order to improve evolution of the solution. Where you have tested ideas experimentally you should present statistics to support decisions about parameters.
4. Evaluation of your solution [20 marks]
You should evaluate your algorithm or algorithms thoroughly. You should present appropriate summary statistics and plots in your notebook to support statements about the effectiveness of the solution, and give appropriate interpretation and discussion of results.
1.5 Marking criteria
2 Evolving Herding Dogs (this part is worth 50 marks)
2.1 Scenario
You need to implement a simple multi-agent environment with two types of agents, sheep (plural) and dogs. All agents are placed on a square, two-dimensional, wrapped around an orthogonal grid made of N × N patches. At the beginning of each run, S sheep and D dogs are placed in the environment in a (pseudo-)random way. Both sheep and dogs can choose from five possible actions: moving in one of the four cardinal directions (North, South, West, East) or staying in the same square. The environment is updated in rounds. All agents are prompted in turn to plan their next move, after which all moves are carried out simultaneously. There is no limit on the number of agents that can reside in the same patch.
Sheep will be using a simple, fixed behaviour, choosing exactly one of the following actions, in order of preference (from highest to lowest):
1. If there is a dog in your current patch, move, if possible, to a patch without a dog;
2. If there is a dog in any of the four adjacent patches (i.e. North, South, East or West of the current one), move, if possible, to an adjacent patch that does not contain a dog;
3. Move to a patch with no sheep, but which is adjacent to a patch with sheep;
4. Move to an adjacent patch containing fewer sheep than the current patch;
5. Make a stochastic choice of action as follows: choose the same action as the last one with 50% probability, or choose one of the remaining four actions, each with 12.5% probability. For the first move, assume for all sheep that their previous move was to stay put.
The initial behaviour for all dogs is to choose one of the five possible actions with the same 20% probability. Dogs can see the position (coordinates of the current patch) of all other dogs and all sheep, and they can tell which species each agent belongs to.
2.2 Task
On the most general level, the goal here is to evolve a team (“pack”) of herding dogs, which would constantly strive to keep all sheep in as tight a formation as possible.
Numerically, we can say the dogs will be aiming to minimise the score defined as the sum of the variances of X and Y coordinates for all sheep:
Design, implement, describe and evaluate a procedure that employs evolution to produce herding behaviour for the dogs that outperforms as best as possible their default behaviour. Each of the dogs should be able to evolve a separate behaviour, which is potentially different from the rest. The size of the environment can be changed at will for evolution purposes, but the resulting behaviour should be tested for 50 sheep, 5 dogs, and an environment of size N = 49, i.e. where patches have coordinates ranging from (-24,-24) to (24,24).
2.3 Reporting
1. Representation of dogs’ behaviour [10 marks]
Choose a clear and efficient representation of your dogs’ behaviour that would also allow for adaptation. Describe the chosen representation and explain the reasons behind it. You can choose a representation where the dog’s behaviour is potentially affected by the value of a function of your choice (think of it as a special sense), e.g. the relative position of the centre of mass of all sheep with respect to your current position, the position of the patch with the highest/lowest number of sheep, the patch with the highest drop in sheep numbers, etc.
Describe clearly how the function is defined, and present a brief a priory argument why it may prove useful.
2. Implementation of default behaviour and fitness estimation [7 marks]
Provide the necessary, working code implementing the simulation where your chosen representation of dogs’ behaviour is set to their default behaviour, then describe how running the simulation is going to provide data for estimating the fitness of your dogs.
3. Design and implementation of adaptation [18 marks]
Describe the design of (9 marks), and implement (9 marks) a procedure that uses adaptation tooptimise the behaviour of your agents.
4. Design of evaluation procedure [10 marks]
Design and describe an evaluation procedure that allows you to compare the behaviour obtained through adaptation to the initial, non-adaptive behaviour, and draw conclusions that are grounded on sound statistical arguments.
5. Experimental evaluation [5 marks]
Collect experimental evidence, carry out, and show the results of the evaluation procedure described above.
2.4 Marking Criteria
1. Three marks for a rational encoding making use of the most relevant information, with up to another 2 marks for a well-argued effort to find a good trade-off in the choice of representation that reduces the search space of distinct behaviours with as little loss of expected performance as possible. Up to another 5 marks for incorporating a well-argued for function providing a potentially useful information to guide a useful herding behaviour.
2. Up to 4 marks for a running, easy to use, and well described implementation of the simulator, and up to 3 marks for the design of a setup that evaluates the dogs’ fitness.
3. A maximum of 9+9 marks for a well implemented, described and argued for a procedure that provides at least the functionality of a generic GA (which would score 3+3 marks max), explores systematically the meta-parameter space to achieve best performance (2+2 marks), and implements at least 3 of the following: fitness scaling or equivalent (e.g. ranking), elitism, niching/crowding, adaptive mutation rates, fitness sharing (up to 4+4 marks).
4. Four marks for a solution that includes an unbiased exploration of the space of initial configurations, six marks for a sound choice and detailed description of a statistical tool/mechanism for the comparison (e.g. choosing the most appropriate statistical test, well argued for).
5. Full (five) marks for a substantial range of experiments, well presented and summarised through the evaluation procedure.
All of the above assumes there is working code implementing the individual objectives. Partial marks may be allocated for design alone. No results will be accepted, if the corresponding parts of the code involved in their generation cannot be executed.
The criteria to be used in the marking also include clarity, simplicity, generality and rigour of the methods chosen, as well as the ability to describe, analyse and visualise experimental results in an effective way.