When we think about how machines make decisions, it often boils down to a few core approaches. One of the most common is optimization, which you can think of as a machine picking the best option from a massive menu. The algorithm searches a list of possibilities and selects an action based on some objective criteria.

Imagine an optimization system trying to find the best route to deliver a package. It could list every possible path and sort them by the shortest distance. Or, it could sort them by the shortest travel time. The distance and duration are the objective criteria—the goals of the optimization. This menu-based logic is a cornerstone of applied artificial intelligence.

Let’s apply this to a simple game like tic-tac-toe. To play like an optimization algorithm, you’d first list all nine squares. Then, you’d cross out the ones that are already taken. Next, you’d need an objective, like choosing the move that leaves you with the most open adjacent squares for future flexibility. This is why the center square is a popular first move—it’s adjacent to eight other squares. After sorting your options by this rule, you pick the top one. If there's a tie, you just choose one at random.

This simple example reveals the first big limitation: optimization algorithms don't actually understand the task. They just follow the sorting criteria. AI pioneer Claude Shannon pointed this out back in 1950 when discussing chess-playing AI. He described two approaches: System A, which programs in actual chess strategies, and System B, which simply searches all possible legal moves. System B is pure optimization—it has no real grasp of chess concepts, it just calculates possibilities.

Searching for the Perfect Solution

Think of optimization algorithms as explorers searching the globe for the highest mountain. The solutions to a problem are like points on a map, and the best solution is the peak of Mt. Everest. If your goal was population density instead of altitude, you’d be looking for Macau.

One way to guarantee you find the highest peak is to visit every single square meter on Earth, measure the altitude, and sort your list. This “brute force search” is impossibly slow for most real-world problems. An industrial process optimization challenge is rarely as small as a tic-tac-toe board.

A more efficient method is to walk the Earth and only take steps in the steepest upward direction. In optimization, this is called a gradient-based method. But this approach has two major flaws. First, where you start your search determines where you end up. If you start in Africa, you’ll climb Mt. Kilimanjaro, not Everest. Once you reach any peak, you can’t go back down to find a higher one. Second, this method only works on smooth, continuous terrain. If you encounter a cliff or a bottomless pit, you can't calculate the slope, and the search fails.

To overcome these issues, more advanced algorithms were developed, many inspired by nature:

• Evolutionary Algorithms: Inspired by natural selection, these create a population of potential solutions, test them, eliminate the ineffective ones, and mutate the survivors to continue exploring.
• Swarm Methods: These mimic how ants or bees explore, using many “explorers” that move across the landscape and communicate their findings to each other.
• Tree Methods: These treat solutions like branches on a tree, similar to a choose-your-own-adventure story. They use smart techniques to search the tree efficiently without visiting every single branch.
• Simulated Annealing: This method is inspired by how metal cools. It starts by exploring a wide area (when the metal is hot) and gradually narrows its search as it gets more confident it’s in the right spot (as the metal cools).

When the Perfect Move Doesn't Exist

Optimization promises to find the “global optima”—the single best possible decision. But in most complex, real-world systems, the idea of a single “perfect move” is an illusion. There are strong moves, weak moves, and creative moves, but rarely a perfect one. That is, unless you’re playing checkers.

In 2007, researchers declared checkers “solved” after nearly 20 years of computation. They proved that perfect play always leads to a draw. But this doesn’t mean we can solve our industrial problems the same way. Checkers is vastly simpler than most real-world challenges, and even then, not every possible position was explored. The complexity of a game is measured by its “branching factor”—the average number of possible moves per turn. For checkers, it's about 3. For chess, it’s 35. For the game Go, it’s 250.

Then there’s uncertainty. In chess, if you move your bishop, it gets there. In a real-life military campaign that chess models, an offensive to take a hill isn't guaranteed to succeed. This uncertainty changes everything.

Scouting Ahead in a Complex World

While we can’t explore the entire future, we can scout a few moves ahead. AI systems like AlphaZero do this by using a tree search to navigate massive possibility spaces like chess and Go. It randomly searches thousands of branches for every move it makes.

This is what gives the AI an “alien playing style.” When it finds a promising path, it pursues it relentlessly, making moves that seem unorthodox or sacrificial to a human player. This is a key area where human-AI collaboration frameworks become essential. Humans don’t search randomly; we use strategy and experience to instinctively ignore millions of nonsensical options. A promising area of research involves using human expertise to bias the AI’s search, guiding it toward more strategically relevant paths.

Following the Manual with Expert Systems

If optimization is a menu of options, an expert system is a detailed instruction manual. It makes decisions by looking up actions from a database of pre-programmed expert rules. This is efficient, but it requires that a human expert already knows how to solve the problem and can write down all the rules.

Consider an HVAC system in an office building. You could create two simple rules: 1) Recycle air to save energy when it’s very hot or cold outside, and 2) Bring in fresh air when CO2 levels get too high. These rules represent the core strategies.

But the real world is messy. How expensive does energy need to be? How much should you close the damper? The answer depends on a nuanced relationship between energy prices, outdoor temperature, and the number of people in the building. A simple two-rule system quickly balloons into hundreds of conditional rules to cover all the exceptions.

This “knowledge acquisition bottleneck” is why early expert systems fell out of favor. It’s incredibly difficult to extract every rule and exception from a human expert and translate it into code. However, these systems are making a comeback in modern autonomous AI system design. Deep inside a self-driving car’s complex learning AI, you’ll find expert rules that take over in safety-critical situations. The learning AI handles the fuzzy, nuanced decisions, while the expert system ensures predictable, safe actions are taken when it matters most.

Ultimately, every method of automated decision-making has its place. The future of applied artificial intelligence isn't about choosing one over another, but about intelligently combining them—using the right tool for the right decision. We need the creative exploration of optimization, the reliable rules of expert systems, and the strategic guidance that defines effective human-AI collaboration frameworks.