5.7 Application Principles of ACO

Construction Graph. A graph can always be constructed using attribute/value pairs as nodes, fully connecting the graph, and placing pheromones on the nodes.
Pheromone Trail Definition. Commonly, τ_i,j can be interpreted as going to node j after node i or the desirability of visiting node i in the jth position.
Exploration vs. Exploitation. The lower pheromone limit and pheromone trail reinitialization of MMAS cater to exploration. An α value greater than 0 caters to exploitation, etc.
Heuristic Information. This information can be static (such as in TSP) or dynamic (such as set covering). Static information only needs to be computed once. Local search often mitigates the need for heuristic information.
Local Search. ACO performs a course-grained search. Local search typically performs a fine-grained search.
Number of Ants. Empirically, multiple ants typically lead to better solutions.
Candidate Lists. Can reduce the search space size.

5.6 Machine Learning Problems

Ant-miner.
Attributes A = {a₁,...}
Values V_i = {v_i1,...}
l (L) Classes B = {b₁,...}
Learn rules of form: IF < term₁ ∧ ... > THEN < b_i >
term has the form < a_i = v _ij >
Ant-miner pseudocode: Figure 5.7.
Construction graph: one node for every possible term plus a dummy start node.
Pheromone trails: τ_ij indicates the desirability of adding (a_i = v_ij) to the current rule.
Heuristic information: based on entropy, but we will ignore in the example that we work.
Solution construction: iterative using equation 5.26.
f(rule) = (TP / (TP +FN)) * (TN / (FP + TN))
Pheromone Update: (1) τ_ij * f(rule) is added to each term of the rule, (2) pheromone values are normalized.
Results: Ant-miner generally has better predictive accuracy than CN2 and returns simpler rules.

Classes: {yes, no}
Attribute: color, Values: {red, blue}
Attribute: size, Values: {small, big}
Training Set: {(yes, red, small), (yes, red, big), (yes, blue, small), (no, blue, big)}