...optimize decisions
Jump to navigation
Jump to search
| AI Solver map |
|---|
|
Your current page is in bold
0 · Start
1 · Predict & analyze
tables, numbers, events
Predict values
Predict categories Forecast time series Cluster / segment Find outliers Make recommendations Find key drivers 2 · Understand unstructured data
text, documents, images, audio
3 · Generate & converse
4 · Decide & act
5 · Build it right
applies to every path
|
YouTube ... Quora ...Google search ...Google News ...Bing News
- AI Solver ... Reinforcement Learning (RL) ... Q Learning ... Deep Q Network (DQN) ... ...forecast time series ... ...make recommendations
Use this page when you must choose actions, not just predict: plan routes, build schedules, set prices, allocate budgets, control machines, pick which offer to show. Many of these problems are solved best by classic optimization, not machine learning.
Start here
- Can you write down the goal and the rules (minimize cost, meet demand, respect capacity and shifts)?
- Yes → mathematical optimization: linear and mixed-integer programming, constraint programming, vehicle-routing solvers (Google OR-Tools, HiGHS, Gurobi, CPLEX; Pyomo or PuLP to write the model). LLMs can help translate a business problem into a model; the solver finds the best answer.
- Shortest path on a known map or graph → classic search: Dijkstra or A*. No learning needed; reinforcement learning only helps when the environment is unknown or changing. Traveling-salesman and delivery routes → routing solvers and heuristics
- Do you repeat the same choice many times and see results quickly (which headline, price, layout, artwork)?
- Yes → A/B test → multi-armed bandits (shift traffic to winners automatically) → contextual bandits (pick the best option for each user or context)
- Is it a sequence of decisions with delayed rewards, and do you have a simulator (games, robots, process control)?
- Yes → Reinforcement Learning (RL): learn a series of actions; Q Learning or Deep Q Network (DQN) for small, discrete action sets; PPO or SAC for continuous control (libraries: Gymnasium, Stable-Baselines3, RLlib)
- Only historical logs, no simulator → offline RL or imitation learning; evaluate carefully with off-policy methods before acting in the real world
- Are the inputs uncertain (demand, travel times)?
- Yes → predict, then optimize: forecast with ranges (...forecast time series), then optimize against them (stochastic or robust optimization)
- Physical robots? → classic control plus perception; for general-purpose manipulation, robot foundation models (vision-language-action models such as Google DeepMind's Gemini Robotics, Physical Intelligence's π0, NVIDIA GR00T), trained largely in simulation and transferred to real hardware
- Training language models with rewards (RLHF, reinforcement learning with verifiable rewards) → ...customize a model
Quick picks
| Problem | Try first | Then compare |
|---|---|---|
| Schedules, shifts, production plans | Mixed-integer programming (OR-Tools, HiGHS) | Constraint programming, heuristics |
| Delivery routes | OR-Tools routing | Commercial routing solvers |
| Shortest path on a map | Dijkstra / A* | — |
| Which offer, price or layout | Multi-armed / contextual bandit | A/B test (simpler, slower) |
| Game, robot or process control with a simulator | PPO, SAC, DQN | Model-based RL, imitation learning |
| Decisions under uncertain demand | Forecast + optimization | Stochastic programming |
Real-world examples
- Delivery routing — UPS's ORION system optimizes each driver's route (more than 200,000 options per route of about 120 stops). UPS expected it to save about 100 million miles, 10 million gallons of fuel and $300–400 million a year once fully deployed. UPS accelerates use of routing optimization software | UPS via 3BL Media
- Data-center cooling — DeepMind's machine-learning system for Google data centers recommended cooling settings and cut the energy used for cooling by up to 40%. DeepMind AI reduces energy used for cooling Google data centers by 40% | Google
- Choosing artwork — Netflix uses contextual bandits to pick which image to show for each title and member, learning from what people actually play. Artwork personalization at Netflix | Netflix TechBlog
Watch out
- Reward hacking — RL agents optimize exactly what you reward, including loopholes; include all the rules and costs.
- Simulator gap — policies that work in simulation can fail on real machines; test gradually with safety limits.
- Feasibility first — a plan that breaks a hard rule (labor law, capacity) is worthless, however cheap.
Measure success
Objective value against the current policy (cost, miles, revenue, energy), constraint violations, regret for bandits, and safety incidents. Compare with a simple rule-based baseline. More: ...evaluate and trust AI
Learn more
- OR-Tools | Google for Developers
- Gymnasium: reinforcement learning environments | Farama Foundation
- Stable-Baselines3: reliable RL implementations
- Vowpal Wabbit: contextual bandits
- Reinforcement Learning (RL) ... Q Learning ... Deep Q Network (DQN)
Where next? ◀ Decide & act · ...forecast time series · ...make recommendations · ...build an agent · ▶ ...evaluate and trust AI