# Embedded Intelligence

**Artificial intelligence is about more than talking computers and robots in search of love and laughter. In fact, AI is most useful in its simplest form: ad hoc decision making capabilities in an embedded system.**

Artificial intelligence (AI) is a field that failed under the weight of its own unrealistic expectations and prophecies. Despite four decades of research and development, the year 2001 came and went with no sign of HAL. Yet intelligence has had a big impact on computing. To use just one example, spell checkers were considered amazingly intelligent applications when they first emerged. Now they're ho-hum.

The truth is that real AI programming is far from glamorous. It basically boils down to solving a problem that assumes some “intelligence” by applying efficient search methods to a directed data set. Intelligent methods can be mixed and matched with more traditional procedural methods-no rule says that the techniques are mutually exclusive. Having a working knowledge of proven AI techniques, and knowing when to use them, enables the programmer to add power, elegance, and sophistication to applications.

**Data representation**

Many search spaces are nicely represented by graphs. So it turns out that an effective way to solve a problem requiring “intelligence” is to put the data in a form easily visualized by a graph. Graphs and graph theory provide visualization aids, formal rules, and techniques that are useful in making ad hoc decisions.

Graphs have their own unique set of terminology. A graph consists of nodes and edges. A node is generally drawn as a circle. An edge is a line between nodes. In software, nodes are usually implemented as data structures with edges as pointers or array indices.

Many algorithms exist to traverse graphs, including depth-first search and breadth-first search. Both depth- and breadth-first searches are blind searching algorithms. The depth-first algorithm searches from some root node to the farthest leaf node along a series of edges, then on to the next leaf node. A breadth-first search, on the other hand, examines all of the nodes one edge distant from the root first, then all of the nodes two edges distant, and so on. (If you aren't familiar with these graph-searching algorithms, you might want to consult an algorithms text like *Introduction to Algorithms* by Thomas H. Cormen or *Computer Algorithms: Introduction to Design and Analysis* by Sara Baase.)

Algorithms like these are “blind” because no information is found during the search that can tell us where the proper solution lies. These searches blindly follow their traversal algorithm, possibly searching every node before locating the answer sought. Knowing when to use a blind search is a pleasant dividend gained from adding AI to your bag of tricks.

**Vehicle navigation**

Let's say we're designing a system that traverses a network of road segments. Maybe it's an automated garbage collection system, traveling camera, or automatic traffic re-routing system. Figure 1 is a map of part of San Francisco.

**Figure 1: A road network**

The first step in creating a graph to represent this data is to determine what elements are nodes. Road intersections are useful, if not obvious, choices for nodes. With those inserted, the picture of the graph is now partially complete. The result is a goal-less, static representation of the map.

The next step is to figure out what additional information the system will need to make intelligent decisions. If the goal of the system is to select the best route for a vehicle to take from one intersection to another, it would be natural to assign weights to the road segments connecting intersections. In the simplest case, there are no one-way streets and all have the same speed limit and number of lanes. (If these things are not true, it's relatively painless to extend our graph and weighting model once we've built it.)

Assigning weights to the graph's edges can help our system find the best route. These weights are determined somewhat arbitrarily-we'll assume they represent the average traffic density. Dynamic weights, based on the time of day or local conditions, are a possibility too, but those don't affect the following analysis.

In Figure 2, the edge weights are designed to represent the average amount of traffic traversing the road per hour (these statistics aren't based on any actual data, but are useful for our discussion). If our vehicle must be routed from the corner of Scott and Jackson (node 5) to the corner of Fillmore and Vallejo (node 17), using the criteria of the least traffic density, our search algorithm should locate the path with the least total weight.

**Figure 2: Traversing the graph**

It's easy to draw our graph on paper, but we still need to represent it within the computer. The two most common representations for graphs are adjacency matrices and adjacency lists.

An adjacency matrix is a static, multidimensional array, with the values in the matrix representing the weights of the edges from one node to another. Figure 3 shows a partial adjacency matrix containing the edge weights between nodes 1 through 6 of our example network. The weight of the edge between nodes 1 and 6 is found in the top right corner (and duplicated at bottom left). The full adjacency matrix for the 20-node road network in Figure 2 would consist of 400 elements, only 31 of which would contain valuable information.

**Figure 3: An adjacency matrix**

An adjacency list is usually implemented as a linked list. Figure 4 shows the adjacency lists for nodes 1 to 6 of our network. The edge weights are not shown in the figure, but are easily stored in the data structure.

**Figure 4: An adjacency list**

When deciding between an adjacency matrix and an adjacency list, you should consider the following factors:

- If the graph is dense or small, represent it with an adjacency matrix. The advantage of a matrix is that the weights can be accessed directly, without the overhead of pointer management and list traversal.
- If the graph is sparse or large, an adjacency list can reduce the amount of wasted memory.
- If new nodes or edges may be added or deleted at runtime, use an adjacency list. Of course, this also means you'll need a dynamic memory capability.

**Heuristics**

If we blindly search the graph following the minimum weight edge out of each node, we may go the wrong way and never wind up at the goal. A more intelligent search can be constructed using heuristics, which are general searching guidelines that work most of the time. Call them rules of thumb. A simple human example is, “It's April and cloudy outside, so take an umbrella.” Just because it's April and cloudy doesn't mean it will rain, but the chances are better than normal under those conditions.

Heuristics are an important part of an efficient, effective search strategy. When in doubt, start by taking an initial stab at the problem-even if it might be off base. For the road traversal problem, one heuristic might be, “When a node has exits on two (or more) equally weighted edges, perform a breadth-first search. Pursue the path with the least total weight.” This helps when node 15 is encountered. Two edges with weight fifteen exit that node. Using breadth-first search, the next level nodes and their exit edge weights are examined. The next level nodes are 14 and 20. Edges with weights fifteen and forty-five exit from those nodes, respectively. Using the least edge-weight criteria, node 14 is selected to continue the search, which is reasonable; node 20 is abandoned.

Some heuristics may sound obvious, but even these can lend insight into the nature of the problem to be solved. The most basic heuristic for the road traversal problem can be stated, “Follow the path with the least edge weights.” That's simple and obvious, but it lays down the baseline rule for the search.

Methods that follow the least edge weight are called greedy algorithms. Greedy algorithms are based on instant gratification. Without looking ahead, a greedy algorithm chooses the cheapest path to follow in the short term. This doesn't guarantee a solution and, on occasion, a less optimal path will be followed. When asked why it followed a particular path, even though it turned out to be wrong, a greedy algorithm would answer, “It seemed like a good idea at the time.”

What seems obvious to a human is not obvious to a machine. Our heuristic, “When a node has exits on two (or more) equally weighted edges, perform a breadth-first search. Pursue the path with the least total weight,” isn't bad, but it has a flaw. What if the least cost path travels away from the goal node? Then we may be choosing a more expensive path. This brings to light the need to know the direction from the current node to the goal node. Developing heuristics in this fashion is a good way of uncovering the bits of core knowledge required to solve a problem.

One way to solve the road network direction problem is to dynamically calculate the distance and bearing between the current node and the goal. This requires latitude and longitude data for each node and floating-point operations every step of the way. This is probably the least optimal solution. A better strategy is to find a set of rules for the differences between pairs of latitudes and longitudes. This helps us distill the knowledge requirements to their minimum set.

The first step is to determine the relationship between direction and latitude and longitude. Figure 5 shows direction derived from latitude and longitude differences.

**Figure 5: Direction, in latitude and longitude terms**

When traveling North, latitude increases; when traveling West, longitude increases; and so on. From these simple relationships, we can eliminate a lot of floating-point operations at each node. Relating this to the map, the starting latitude and longitude at the intersection of Scott and Jackson is N37 47.514 latitude and W122 26.356 longitude. The goal intersection at Fillmore and Vallejo is N37 47.725 latitude and W122 26.002 longitude. Applying the compass rules from Figure 5, it's apparent that the goal intersection is northeast of the starting intersection.

Six individual rules are built from the direction relationship:

Rule 1: if Lat(goal) > Lat(current) then goal is North

Rule 2: if Lat(goal) < lat(current)="" then="" goal="" is="">

Rule 3: if Lat(goal) = Lat(current) then goal is Null

Rule 4: if Lon(goal) > Lon(current) then goal is West

Rule 5: if Lon(goal) < lon(current)="" then="" goal="" is="">

Rule 6: if Lon(goal) = Lon(current) then goal is Null

We can combine rules from this basic set to accommodate complex directions, such as Northeast and Southwest. This is done simply by ANDing the basic rules together. The legal rule combinations are:

Rule 1 ^ Rule 4 -> goal is Northwest

Rule 1 ^ Rule 5 -> goal is Northeast

Rule 1 ^ Rule 6 -> goal is North

Rule 2 ^ Rule 4 -> goal is Southwest

Rule 2 ^ Rule 5 -> goal is Southeast

Rule 2 ^ Rule 6 -> goal is South

Rule 3 ^ Rule 4 -> goal is West

Rule 3 ^ Rule 5 -> goal is East

Rule 3 ^ Rule 6 -> goal is Null

Including the basic and complex rules results in a successful search method. If the goal is Northwest of the current node, North and East moves are legal, but South and West moves are not.

When the search arrives at node 12, the logical path to take is the edge from node 12 to node 11 with the weight of 15. The direction is North, which appears to be legal, and the edge weight is minimal. This is entirely wrong, since the search steers away from the goal node. Now is when the our rule base should constrain the search. Node 12 is parallel to node 17, so rule 3 is true. The longitude is decreasing, so rule 5 is true. If rule 3 and rule 5 are true, the goal is due East. The rule base has done its job-overriding and guiding the otherwise blind search. This is shown in Figure 6.

**Figure 6: Rule base violation**

**The blind leading the blind**

As we've seen, graph representations and blind searching algorithms alone are not sufficient to solve most problems. But combining these techniques with heuristics and sets of problem-specific rules, as we have done here, results in a valuable use of artificial intelligence. Similar techniques can be applied in a variety of application domains. Although our example concentrated on static data, the techniques shown here are perhaps most useful when edges and edge weights are changing and no hard coding of rules is possible.

Obviously, embedded systems suffer some unique constraints. Recursion is usually not allowed in embedded programming, though it is a common technique in graph-searching algorithms. Dynamic memory allocation is generally frowned upon in embedded systems for critical applications, but it's difficult to add and delete nodes from linked list data representations without it. In light of these considerations, here are some tips for embedding intelligence:

- Consider offloading parts of the processing to a more capable system. Perhaps the embedded system need only solve part of the problem on the fly.
- Avoid recursion. Any recursive function can be rewritten in an iterative function.
- Keep dynamic memory allocation to a minimum. If a list's size remains fairly constant, use an array instead. Size the array for the maximum size of the list and return failure if there's an attempt to exceed that maximum.
- Think of intelligence in terms of an insect instead of a supercomputer. Think about how low-level life forms deal with unexpected situations and obstacles.

Above all, use combinations of blind, greedy, and intelligent searching algorithms. There's certainly no restriction that only one method must be used to solve a problem requiring intelligence.

**Jeff Stefan** is a lead technical development engineer at OnStar. He holds a BS in computer science from Lawrence Technological University and is a frequent speaker at the Embedded Systems Conference. His e-mail address is .

**AI References**

Dean, Thomas, et. al. *Artificial Intelligence Theory and Practice.* Reading, MA: Addison-Wesley, 1995. This is a good text with a wide range of subject coverage. As with most AI books, all of the examples are in Lisp.

Charniak, Eugene and Drew McDermott. *Introduction to Artificial Intelligence.* Reading, MA: Addison-Wesley, 1985. Although this is an older book, it's filled with practical examples. This is another Lisp-centric text.

Bratko, Ivan. *Prolog Programming for Artificial Intelligence, Third Edition.* Reading, MA: Addison-Wesley, 2000. This book expresses AI concepts and algorithms using PROLOG, another popular AI programming language. The book is constructed in two parts, the first part covers PROLOG programming and the second part covers AI concepts and algorithms. The sections on searching are excellent.

Winston, Patrick Henry. *Artificial Intelligence, Third Edition.* Reading, MA: Addison-Wesley, 1992. This is an AI classic, loaded with detailed pseudocode.

**Algorithm References**

Cormen, Thomas H. et. al. *Introduction to Algorithms.* Cambridge, MA: MIT Press, 1997. This is a definitive reference, providing consistent pseudocode throughout the text.

Basse, Sara and Allen Van Gelder. *Computer Algorithms, Third Edition.* Reading, MA: Addison-Wesley, 2000. This excellent book provides Java and Java-like pseudocode for all of the examples and concepts.

Return to the May Table of Contents