The Dynamic Analysis and Replanning Tool (DART) is an artificial intelligence program used by the U.S. military to optimize and schedule the transportation of supplies or personnel and solve other logistical problems. In the first 4 years of its use, DART saved the military enough money to recoup the previous 30 years of investments.
DART is one of many examples of utilizing AI for planning. Airports use AI for planning flights; Google Maps and other navigation systems use AI for planning; the Mars Rover uses AI to navigate the Red Planet; many large corporations utilize AI planning algorithms to plan logistics and personnel.
Using a simplified entrepreneurial analogy, this article will explain a planning approach called “Serial Planning Graphs.” In our example, an entrepreneur wants to take an idea stirring in his head and create an app…he wants to go from idea to delivery. The graph below shows the steps for our example (and will be explained in detail). Those who are familiar with entrepreneurship know that this is a contrived example, and there are indeed dozens to hundreds of additional scenarios and steps. Those who are familiar with A.I. also know that this is a contrived example, and there are indeed hundreds to thousands to millions of additional states and actions.
Here are some key AI concepts used with Serial Planning Graphs:
- In AI, a scenario is called a “state,” which means any condition or scenario that an agent can be in (an agent being the actor, like an entrepreneur or her company). A step, next-step, or action-item in AI terminology is simply called an “action.” So, all AI planning graphs are comprised of states and action.
- When moving from a state to a new state, an action is required. As with anything in life, there is a cost for an action. This may be a monetary cost, an opportunity cost, or an emotional cost, depending on what you are planning. For example, the cost of seeing a movie is associated with a monetary cost of the price paid for admission, and the opportunity cost of missing a hike with friends. The cost of creating an app is both a monetary cost, and possibly an opportunity cost of focusing on other endeavors.
- Finally, when given several choices for actions, or several states resulting from a single action, the actions and states can be in “mutex” with each other, which simply means that you cannot choose two actions that are mutually exclusive, or be in two states that are mutually exclusive. For example, the action of either “driving to the movies” or “walking to the movies” are mutex to each other; and the states of “being in the movie theater” and “being on a mountain” are mutex to each other. For most entrepreneurs, especially startups, the company’s focus creates a mutex with other areas of focus.
As humans, we can process dozens of states and actions in a planning graph. It takes us a long time and is prone to error and fatigue. A computer can process millions of states and actions with accuracy and precision—and minimal time.
In relation to the example and serial planning graph of the journey of an entrepreneur’s idea to an app, the first thing to notice is the “cost” at each “level.” There is always a cost with every action in a plan. Most of the time, the best plan involves the least cost. For example, the least cost navigation plan is the route that gets you from your house to the movies the fastest (and as we’ll see in our example, the least cost app development plan is to pay a developer cash, or an equilibrium hybrid of cash and equity). Many entrepreneurs will start with an idea, an end goal, and plow forward without a plan. While the persistence and gusto are admirable, the lack of a plan can lead to failure (for entrepreneurs, reported failure). It’s a wise idea, then, to examine the various paths to a goal, and the costs associated to each action at each level.
The value of an action is determined by a “heuristic,” which means a way of assessing the cost to reach the next level, or the cost of reaching the end goal, from the current level and state. In our example, we created a rather simple heuristic of “add 1” at each level-action pair, except for the development of the product. For the development of the product, we assigned a cost of 15 if the developer is working nights and weekends, and a cost of 5 if the developer is working full-time.
In our example, we can see that the cost to launch a product will be either 18 or 7. This is obvious when all states and actions are fully mapped in advanced. However, life doesn’t work that way. Our example is called “fully observable,” yet life, love, and business are usually “partially observable” (at best). As such, the need for an accurate heuristic is vital. In other words, in a partially observable game (the term “game” is often applied to many search and planning algorithms in AI, not just video or board games), an accurate heuristic or “cost per level” can help guide us to choose the next best action. This usually starts with the question: “Can I get to my end goal from where I am now?”
This is such a vital precondition of any heuristic: The cost value of a heuristic needs to assess how likely it is that an agent can reach an end goal from the current state. The likelihood that an agent can reach an end goal from its current state was the essence of the IBM Deep Blue algorithm that beat world chess champion Garry Kasparov in 1996. The algorithm, immersed in a partially observable game, ran a heuristic that continually assessed the likelihood of winning the game based on: (1) the current state of the game; (2) actions that are possible from this state; and (3) the resulting state of the game after the action. This strategy is the essence of artificial intelligence, and all intelligent planning. It is also an example of “forward searching.”
As its name implies, “forward searching” starts at the current state you’re at, and makes “next action” choices based on some heuristic. But there’s no reason why your initial state cannot be your final goal, and your final goal your initial state! This is called “backward searching” (of course!), and is sometimes a very powerful search tool.
In our example of app creation, we would start from the state of “Launch Product,” and work our way backward to “idea”, using a heuristic to choose our actions. Once we’ve reached our end goal (initial state), we unwind the steps to create our plan.
This has been a 100,000-foot overview of AI serial planning graphs. While high-level, the major concepts have been introduced. AI algorithms, in conjunction with humans, can create powerful search and planning algorithms using intelligent heuristics to solve problems ranging from military troop deployment to Amazon delivery logistics, and entrepreneurial planning.
Learn more about AI and how it can help your business by taking the first step: contact us.