Management and Tech Notes

Mgmt | Tech | Productivity posts.
Catch-all of random thoughts.

[Management and Tech Notes @padlogk]

Branch-and-cut solves a relaxed puzzle, spots how the answer cheats, and adds a rule that bans just that cheat. When cutting stalls, it branches. That's how big TSPs get solved. #Algorithms #TSP Level 2 · ELI10

[Management and Tech Notes @padlogk]

Your quick answer cheats by drawing little separate loops? Add a rule: no little loops! 🔪 Branch-and-cut keeps slicing away cheats until the perfect trip remains. #Algorithms Level 1 · ELI5

[Management and Tech Notes @padlogk]

Branch-and-bound is the engine inside Concorde, CPLEX, Gurobi and SCIP: no better worst case, but proven optimality and an anytime gap. Add cuts or columns and it scales. #Algorithms #TSP Level 6 · Researcher

[Management and Tech Notes @padlogk]

Why branch-and-bound is exact: valid lower bounds + branches that partition the feasible set ⇒ nothing better is ever pruned. Bounds come from relaxations. #Algorithms #TSP Level 5 · Graduate

[Management and Tech Notes @padlogk]

Branch-and-bound = branching rule + optimistic bound + search order. Prune when lb ≥ incumbent. Still exponential in the worst case, but exact and fast in practice. #Algorithms #TSP Level 4 · Undergrad CS

[Management and Tech Notes @padlogk]

Branch-and-bound on 4 cities: seed a tour of 80, then one bound check (95 > 80) wipes out a whole subtree. 80 is proven optimal, not just found. #Algorithms #TSP Level 3 · High-school

[Management and Tech Notes @padlogk]

Branch-and-bound asks one question of every half-built route: "Even in the best case, could you beat my champion?" If not, the whole branch gets cut. #Algorithms #TSP Level 2 · ELI10

[Management and Tech Notes @padlogk]

Half-built route already longer than your best full route? Stop right there ✂️ Branch-and-bound skips hopeless routes and still finds the perfect answer. #Algorithms Level 1 · ELI5

[Management and Tech Notes @padlogk]

ACO is the model-based metaheuristic: pheromone defines a sampling distribution, a precursor of learned constructive policies. Competitive only with 2-opt/LK local search. #Algorithms #TSP Level 6 · Researcher

[Management and Tech Notes @padlogk]

ACO is model-based search: pheromone τ is a distribution over solutions, refined by reinforcement. MMAS (τ clamped to [τmin, τmax]) and ACS beat the original Ant System. #Algorithms #TSP Level 5 · Graduate

[Management and Tech Notes @padlogk]

ACO: ants choose j with p ∝ τ^α·η^β (η = 1/d), then pheromone updates as τ ← (1−ρ)τ + Σ Q/L_k. Each iteration costs O(m·n²); ACS adds q0 exploitation. #Algorithms #TSP Level 4 · Undergrad CS

[Management and Tech Notes @padlogk]

ACO on 4 cities: from A an ant picks B 59% of the time (1/d² weights) and builds the 80-tour. After evaporating (ρ=0.5) and depositing 100/L, A–C and B–D lead at 2.80. #Algorithms #TSP Level 3 · High-school

[Management and Tech Notes @padlogk]

In ant colony optimization, each ant builds a tour step by step, biased toward edges with high pheromone and short length. Trails evaporate, and short tours deposit more. #Algorithms #TSP Level 2 · ELI10

[Management and Tech Notes @padlogk]

Ants find short paths with no map 🐜: short trails get more smell, long ones fade, and everyone follows the strongest. Ant colony optimization does this for tours. #Algorithms Level 1 · ELI5

[Management and Tech Notes @padlogk]

Game-tree search lineage: minimax → alpha-beta → Deep Blue, then two branches: Stockfish+NNUE (αβ + learned eval) and AlphaZero/MuZero (MCTS + self-play). They're converging. #Algorithms #AI Level 6 · Researcher

[Management and Tech Notes @padlogk]

AlphaZero: one network gives (p, v); MCTS selects by PUCT, Q + c·P·√ΣN/(1+N), improving the policy. Training minimizes (z−v)² − πᵀlog p + c‖θ‖². #Algorithms #AI Level 5 · Graduate

[Management and Tech Notes @padlogk]

Minimax backs up max/min values to depth d with an eval at the cutoff: O(b^d). Alpha-beta prunes branches that can't matter, reaching O(b^(d/2)) with good move ordering. #Algorithms #AI Level 4 · Undergrad CS

[Management and Tech Notes @padlogk]

The poisoned pawn: grabbing it scores +1 or −8 (queen fork), so MIN makes it −8. A quiet move scores +2 or +3, so MIN makes it +2. Minimax plays quiet; greedy loses the queen. #Algorithms #AI Level 3 · High-school

[Management and Tech Notes @padlogk]

From greedy to AlphaZero: minimax imagines the opponent's best reply, alpha-beta skips useless branches, and self-play teaches the program to judge positions itself. #Algorithms #AI Level 2 · ELI10

[Management and Tech Notes @padlogk]

Grabbing a piece feels good until your friend strikes back ♟️ Good game robots imagine future turns, pick the move that wins, and learn from every game. #Algorithms Level 1 · ELI5