Clustering, association rules, reinforcement learning and genetic algorithms
Use four different learning/optimization mechanisms, then apply adversarial-learning ideas to the GAN challenge.
What you need to be able to do
- describe clustering from feature similarity;
- describe association-rule discovery and interpret a scenario;
- describe RL states, actions, rewards, policies, cumulative reward and exploration/exploitation;
- describe GA population, fitness, selection, crossover, mutation, evaluation and termination.
Rapid Recall Deck
Attempt each card first. If a term is new, flip and study it, then answer it again without looking. Mark only secure recall as “Know it”.
Unsupervised patterns
Clustering
A shop has spend, visits and discount-use features but no segment labels. What can clustering do?
Group customers with similar feature patterns into unlabeled clusters that analysts can later interpret as useful segments.
Association rule
High vandalism areas also often show theft. What has been discovered—and what has not?
An association/co-occurrence relationship has been discovered. It can support resource allocation/investigation but does not by itself prove vandalism causes theft.
Reinforcement learning interaction
state → choose action using policy → environment ↑ ↓ └──────── reward + next state ──┘
Robot navigation
+10 for goal, −5 collision, −0.1 per step. Explain state, action, reward, policy and cumulative reward.
State represents the robot's current situation; actions are movement choices; rewards evaluate consequences; the policy selects actions from states; learning aims to maximize cumulative long-term reward across the route.
Explore vs exploit
Why explore?
Why try an action not currently believed best?
Exploration gathers evidence about less-known actions and can discover a better policy that pure exploitation would miss.
Why exploit?
Why not explore randomly forever?
The agent must use what it has learned to gain reward; endless exploration wastes performance on uncertain/poor actions.
Genetic algorithm cycle
population → fitness evaluation → selection
↓
crossover + mutation
↓
new population
↓
repeat until terminationTravelling-salesperson route
Explain what a candidate, fitness function, selection, crossover, mutation, evaluation and termination could mean.
A candidate encodes a route; fitness scores route quality such as distance; selection favours useful parents; crossover combines route information; mutation introduces variation; new routes are evaluated; iteration stops at a chosen fitness/convergence/generation condition.
Match problem to learning signal
Four goals
Customer groups without labels; products bought together; robot navigation from rewards; optimize a route.
Clustering; association rules; reinforcement learning; genetic algorithm.
GAN generator/discriminator balance
Explain the challenge
In the Visionary Studios GAN, a generator turns a D-dimensional noise vector into synthetic images while a discriminator judges real versus generated images. Explain how the adversarial dynamic works and why imbalance can cause training instability or mode collapse.
The generator learns to fool the discriminator; the discriminator learns to distinguish real from fake. Their competition supplies the learning signal. If one becomes too dominant, useful feedback can weaken and training can become unstable. Mode collapse occurs when the generator produces a narrow range of similar outputs instead of dataset diversity.
Mechanism → suitability
Paper 1 mini-set
1. Describe exploration vs exploitation. 2. Describe selection, crossover and mutation. 3. Justify RL rather than supervised learning for sequential navigation learned from rewards.
Exploration learns uncertain actions; exploitation uses current knowledge. Selection chooses parents, crossover recombines, mutation varies. RL fits navigation because the agent learns from sequential state-action-reward interaction rather than receiving a correct labelled action for every state.
Finish the learning cycle
Case study / exam preparation
Deepen the GAN research note with generator/discriminator, adversarial dynamic, D-dimensional noise vector, instability and mode collapse.
Paper 1 — short practice
Reconstruct clustering, association, RL and every GA stage from memory.
Case-study checkpoint
Add one credible GAN-stabilization/mode-collapse mitigation source and explain why it could help.