COMP90054 AI Planning for Autonomy
The Complete Exam Bible · SM2 2026COMP90054 Overview
- Faculty of Engineering and Information Technology
- Semester 2, 2026
- a graduate coursework subject
- 12.5 points
- an advanced AI planning and reinforcement-learning subject
COMP90054 develops autonomous-agent modelling, search and heuristics, classical planning, Markov decision processes, reinforcement learning, game theory and the ethical evaluation of agents that reason about action. It is taught within Faculty of Engineering and Information Technology. It is a graduate coursework subject.
- Representation is the first algorithm State, action, goal and reward choices determine which solutions are even expressible.
- A heuristic changes work, not truth Search guidance must preserve the guarantees claimed for the chosen algorithm.
- Return is not responsibility A high learned reward can still reflect an unsafe proxy or excluded stakeholder.
- Current S2 facts are first-party SM1 material supports stable concepts only; S2 assessment and hurdle facts come solely from the 2026 Handbook.
How COMP90054 is assessed
| Component | Weight | Format |
|---|---|---|
| Individual Programming Assignment | 25% | Due in Week 6; approximately 30 hours of work |
| Online Quiz | 10% | 30 minutes in Week 7 |
| Individual Modelling Assignment | 15% | Due from Week 11 to Week 12; approximately 15 hours of work |
| Closed-book Final Examination · hurdle | 50% | Two hours during the official examination period; must be passed |
Current Semester 2 assessment is a 25% individual programming assignment in Week 6, a 10% 30-minute online quiz in Week 7, a 15% individual modelling assignment in Weeks 11-12, and a 50% two-hour closed-book final exam. The examination is a hurdle and must be passed to pass the subject.
Current COMP90054 dates
| Date | Item | Control |
|---|---|---|
| Week 6, Semester 2 2026 | Individual Programming Assignment | 25%, approximately 30 hours of work. |
| Week 7, Semester 2 2026 | Online Quiz | 10%, 30 minutes. |
| Weeks 11-12, Semester 2 2026 | Individual Modelling Assignment | 15%, approximately 15 hours of work. |
| Official examination period | Closed-book Final Examination | 50%, two hours, pass hurdle. |
Current-offering dates captured in the course materials. Confirm changes and exact submission settings in the live LMS.
What COMP90054 covers
Build the course in three arcs: Autonomous Agents, Modelling and Classical Planning establishes the frame, MCTS Selection and Backup deepens it, and Planning, Multi-Agent Reasoning and Ethical Control tests the complete method.
Autonomous Agents, Modelling and Classical Planning
autonomous agent · state · planning model · translate a real task into states, actions, transition effects and goals02State-Space Search and Path Cost
search node · frontier · path cost · compare breadth-first, depth-first and uniform-cost search by frontier rule and cost assumptions03Heuristic Search and A-star
heuristic · admissibility · consistency · apply A-star and audit the heuristic needed for the claimed optimality04Delete Relaxation and Planning Heuristics
delete relaxation · relaxed plan · helpful action · construct a relaxed planning graph and use it to guide classical search05Markov Decision Processes and Bellman Reasoning
Markov decision process · policy · Bellman optimality · calculate state or action values and select a policy under uncertainty06MCTS Selection and Backup
Monte Carlo tree search · UCT exploration bonus · backup · apply UCT selection and interpret estimates under a limited simulation budget07Bandits and Monte Carlo Prediction
multi-armed bandit · epsilon-greedy · Monte Carlo return · estimate action or state value from sampled returns and choose an exploration strategy08SARSA and Q-Learning Control
temporal-difference learning · SARSA · Q-learning · calculate and compare on-policy and off-policy action-value updates09Policy Gradient and Actor-Critic
policy gradient · actor · critic · derive the direction of a policy-gradient or actor-critic update and interpret its estimator10Deep Tree Search and Learned Guidance
policy prior · value network · self-play · explain how learned policy and value estimates guide tree search and training11Planning, Multi-Agent Reasoning and Ethical Control
multi-agent system · best response · ethical consequence · connect strategic interaction or advanced planning to explicit human and stakeholder controlsIt carries 12.5 points.
It is positioned as an advanced AI planning and reinforcement-learning subject.
The subject joins symbolic planning and learning methods in one policy-selection problem, with a final-exam hurdle and two individual technical assignments requiring both implementation and modelling judgment.
Assessment in comp90054 is distributed as follows: a 25% programming assignment, 10% online quiz, 15% modelling assignment and 50% closed-book final examination
The operational assessment conditions matter here.
The current final is a two-hour closed-book examination during the official exam period and is a pass hurdle.
What makes comp90054 demanding is concrete: choosing the correct state and objective representation before applying an algorithm, then distinguishing convergence or search performance from the safety and ethical quality of the resulting policy
The 50% examination is a hurdle and must be passed to pass COMP90054; the overall subject result must also satisfy the University pass standard.
For enrolment planning, The current overview assumes confident Python and provides no Python teaching; students must confirm the formal eligibility rules in the live Handbook.
Build the course in three arcs: Autonomous Agents, Modelling and Classical Planning establishes the frame, MCTS Selection and Backup deepens it, and Planning, Multi-Agent Reasoning and Ethical Control tests the complete method.
Choose a planning method for a delivery agent
- 1Define state, action, goal and safety constraints.
- 1Separate known deterministic planning from uncertain execution.
- 1Choose a baseline search or policy method and explain its assumptions.
- 1Define performance, compute and failure metrics.
- 1Test blocked, unreachable and unsafe states before recommending deployment.
Key terms
- autonomous agent
- Active system that perceives, reasons and acts toward goals or rewards in an environment. This chapter uses the concept when students translate a real task into states, actions, transition effects and goals.
- state
- Representation containing the information treated as relevant to future action choice. It helps explain the reasoning required to translate a real task into states, actions, transition effects and goals.
- planning model
- Formal description of actions, preconditions, effects, initial conditions and goals. Its limit matters because omitted variables, incorrect observability and abstraction choices can make a formally solved model unsafe in reality.
- search node
- Record representing a state together with path, cost and frontier information used by a search algorithm. This chapter uses the concept when students compare breadth-first, depth-first and uniform-cost search by frontier rule and cost assumptions.
- frontier
- Set of generated nodes awaiting expansion. It helps explain the reasoning required to compare breadth-first, depth-first and uniform-cost search by frontier rule and cost assumptions.
- path cost
- Accumulated cost of actions from the initial state to a node. Its limit matters because completeness and optimality claims depend on branching, positive costs and duplicate handling.
- heuristic
- Estimate of remaining cost or distance from a state to a goal. This chapter uses the concept when students apply A-star and audit the heuristic needed for the claimed optimality.
- admissibility
- Property that a heuristic never overestimates true remaining optimal cost. It helps explain the reasoning required to apply A-star and audit the heuristic needed for the claimed optimality.
- consistency
- Property requiring heuristic change across an edge not to exceed that edge cost. Its limit matters because an informative heuristic can still be invalid, expensive or misleading under a changed state representation.
- delete relaxation
- Planning relaxation that ignores action delete effects, making facts persist once achieved. This chapter uses the concept when students construct a relaxed planning graph and use it to guide classical search.
- relaxed plan
- Action sequence solving the relaxed problem rather than the full problem. It helps explain the reasoning required to construct a relaxed planning graph and use it to guide classical search.
- helpful action
- Action selected from an early relaxed-plan layer as potentially useful search guidance. Its limit matters because ignoring deletes can hide resource conflict and subgoal interaction, so a low estimate is not an executable plan.
- Markov decision process
- Model of sequential decisions with states, actions, transition probabilities, rewards and discounting. This chapter uses the concept when students calculate state or action values and select a policy under uncertainty.
- policy
- Rule mapping states or observations to action choices. It helps explain the reasoning required to calculate state or action values and select a policy under uncertainty.
COMP90054 FAQ
How is comp90054 assessed?
a 25% programming assignment, 10% online quiz, 15% modelling assignment and 50% closed-book final examination
What is the comp90054 exam or final-task format?
The current final is a two-hour closed-book examination during the official exam period and is a pass hurdle.
Does comp90054 have a hurdle or component-level pass rule?
The 50% examination is a hurdle and must be passed to pass COMP90054; the overall subject result must also satisfy the University pass standard.
Which offering does this comp90054 guide cover?
It is aligned to Semester 2, 2026; confirm your enrolled class and timetable in the current institutional system.
Is this comp90054 resource an official university guide?
No. It is an independent comp90054 study resource; current institutional instructions remain authoritative for assessment operation.
How should autonomous agent be used in COMP90054?
Active system that perceives, reasons and acts toward goals or rewards in an environment. This chapter uses the concept when students translate a real task into states, actions, transition effects and goals. Use it to translate a real task into states, actions, transition effects and goals; remember that omitted variables, incorrect observability and abstraction choices can make a formally solved model unsafe in reality.
Where do students usually lose marks in comp90054?
choosing the correct state and objective representation before applying an algorithm, then distinguishing convergence or search performance from the safety and ethical quality of the resulting policy
What prerequisites or restrictions apply to comp90054?
The current overview assumes confident Python and provides no Python teaching; students must confirm the formal eligibility rules in the live Handbook.
How to study for the exam
Retrieve the course map, practise the recurring method—formalise the autonomous agent's states, actions, transition or observation model and objective, apply a search, planning or learning update, compare policies against a baseline, then evaluate computational, safety and stakeholder consequences—on changed scenarios, and verify every operational assessment detail in the live institutional system.
Your AI Artificial Intelligence tutor for COMP90054
Stuck on a hard COMP90054 question? Sia is AskSia’s AI Artificial Intelligence tutor — ask any COMP90054 AI Planning for Autonomy question and get a clear, step-by-step explanation grounded in how the course is actually taught and assessed. Read this whole study guide free, then take your hardest questions to Sia.
