SEMTM0016 Artificial Intelligence for Robotics
SEMTM0016 Coursework - Part C
Task Overview
In this part of the coursework, you are being assessed on your ability to:
• Discuss and explain for which problems/tasks an approach and algorithm is suitable and why.
• Implement and apply a range of algorithms to a given “real-world” scenario, including poten- tially noisy, unstructured or unreliable datasets.
• Evaluate and appraise the performance of a range of algorithms.
Task Description
Machine learning in robotics has many different approaches and algorithms available which can be applied to various problems. Therefore, it is a key skill to be able to select the appropriate approach and algorithm, to configure it appropriately, and to evaluate the performance. This assessment is an opportunity for you to demonstrate this capability.
You are the mighty HeroBot traversing the MazeDungeon where you could encounter many differ- ent entitities. Herobot’s overall goal is to navigate the MazeDungeon and reach the exit as efficiently as possible. However, to achieve this overall goal there are various tasks Herobot will need to do, such as distinguishing between friend and foe or deciding what action to take upon meeting an enemy in the maze.
The following will help you in your quest and can be downloaded from Blackboard or via the GitHub link:
• dungeon images colour80. zip : This dataset is comprised of 17,842 80x80 pixel RGB colour images of entities that could be found within the dungeon. Each entity is labelled with a particular ‘race’ and there are five races in total.
• dungeon sensorstats partC . csv : This dataset gives sensor and stats information about 10,000 example entities that may reside within the dungeon.
• DungeonMazeWorld: The base reinforcement learning environment for the maze solving grid- world problem.
Given the datasets and base simulation environment provided, in this part of the coursework you are now expected to identify and solve three tasks demonstrating each of the three machine learning approaches (i.e. one task per approach);
• a supervised learning approach,
• an unsupervised learning approach, and,
• a reinforcement learning approach.
The tasks cannot be exactly the same as those covered in Part A and B, but must be different in some way. For example, for a supervised learning task you might use classification on the image dataset again but no longer restricted to two classes. If you have any doubts or questions, please contact a member of the teaching staff.
You should look to evidence and demonstrate your critical thinking through the written commu- nication of your approaches. You must produce a report which effectively communicates, for each task:
• a description of the proposed task, your choice of algorithm to solve it and why that algorithm is appropriate.
• a description of how you implemented the algorithm, including any pre-processing performed on the datasets and any adaptations made to the base simulation environment.
• justifications for any design choices, which should be supported with evidence such as tables and plots where appropriate.
• an evaluation and critical appraisal of the performance of the algorithm, using suitable metrics and comparisons to baselines where appropriate.
Report
Your report should be up to but no longer than eight pages, 11 or 12pt font is sufficient.
Your report must be submitted as a pdf and should be prepared either in LaTeX (overleaf is a good approach), MS Word, or a similar text editor to prepare the report and submit it as a pdf document.
Your code will not be marked for elegance, but it should run correctly; it is expected you will use Python. Do not include screenshots of graphs, they should be imported directly; resize them to the correct size before importing them, if the labels are tiny the graphs will not be marked. Make sure figure captions are descriptive, it is better to have some overlap between figure captions and the main text than to have figure captions that are not reasonably self-contained.
Avoid code snippets in the report unless that feels like the best way to illustrate some subtle aspect of an algorithm; do always though consider a mathematical description if possible. You will be asked to submit your code and it will be tested to make sure it works and matches your report. It will not, however, be marked itself for quality.
Assessment Criteria
Your report will be assessed with consideration to the following unit criteria and general University marking criteria and scales. Feedback will be provided addresing the same criteria where appropri- ate.
|
Criteria
|
Weight
|
|
Task and algorithm choice
- Identify and describe three relevant tasks.
- Realistic and challenging scope of tasks.
- Approach and algorithm choice is suitable for each task.
- Suitable justification given for choice of approach and algorithm.
|
0.2
|
|
Implementation
- All algorithms implemented correctly.
- Sufficient level of detail in documentation for reproducibility of work.
|
0.2
|
|
Design choices
- Evidence of critical thinking/reasoning.
- All design choices are fully discussed and justified.
- Appropriate methods used to select any hyperparameters where applicable.
- Justifications are supported by suitable evidence where appropriate.
|
0.25
|
|
Evaluation and self-reflection
- Choices of evaluation metrics/method are appropriate, discussed, and justi- fied.
- Evaluation includes comparison to suitable baselines where appropriate.
- Work completed is critically appraised.
- Limitations or shortcomings in relation to original task are acknowledged and discussed.
- Realistic recommendations given for future work.
|
0.25
|
|
Report presentation
- Report has a logical structure and clarity of presentation.
- Excellent quality of writing, spelling, grammar.
- Tables, diagrams and figures are readible, captioned and referenced in main body of text.
- Use of English is clear and readible.
- Communication is appropriate for audience.
|
0.1
|