• Contribute
• What's New
• Submit News

Steve did his undergraduate work at Trinity University in mathematics and graduate study at MIT in aeronautical and astronautical engineering. He spent about 14 years working for the U.S. Army in various aviation related programs including Aircraft Survivability Equipment and the Apache program. In the Apache program he was the Army's chief engineer for the TADS/PNVS (Target Acquisition and Designation Sight / Pilot's Night Vision Sensor) during the end of the development phase and the beginning of production. In 1984 he founded System Dynamics International's (SDI) St. Louis Technical Center and continued work on mostly Army aviation and weapons systems projects. In 1988 he wrote some Pascal code to implement what he called 'evolutionary problem solving' (he wasn't aware of the other work that had been done prior to this in genetic algorithms). This code was the kernel from which SDIs evolutionary algorithm, e, grew. SDI subsequently has applied e to various target detection and classification problems as part of its government research and had good enough success that they thought e might make a good commercial product. Along with Dr. Daniel Dunay, he collaborated with several others to develop a commercial version of e, which they began to sell a couple of years ago. Thank you to Steve for the interview!

G5: What kind of AI-related projects has SDI been involved in recently?

Right now we are working on a project with another company to identify target chips from a Synthetic Aperture Radar (SAR). This type of radar can provide fairly high resolution images of a swath of terrain. Detection software operates on this image to find candidate targets. Those portions of the image extracted for further examination are called chips. Our job is to examine those chips and identify the type of target ( e.g. M1 tank, T80 tank, M35 truck, etc.) A couple of approaches are taken to make this identification. Both approaches first require operating on the image to turn it into a line drawing or a set of points. Based on this processed image the first approach computes a set of features (for example, the ratio of the circumference to the area of the object). Usually 10 to 20 features are used. The second approach compares the live image chip (processed into points) to a set of models and computes a set of features that measure how well the live chip matches the model chip (for example, the percent of points that match). If we are trying to determine if the live chip is one of 10 different target types then we need to have models of each of these target types at lots of different aspects. When all these numbers are computed we turn the information over to our evolutionary algorithm, e, to develop a computer program that will identify the target type. This is the only part of the job that falls under the AI umbrella…the rest is pretty standard stuff except that a unique approach is used to do the model comparison.

G5: Could you explain how evolutionary algorithms could be used in target detection and acquisition?

As described In the problem above, once we have a set of input features defined, we use the evolutionary algorithm to come up with a classification program that uses a set of features as inputs, and outputs a classification code. In the simplest approach the evolutionary algorithm runs off line in the laboratory, and only the resulting program would actually be used on line in the target detection or classification problem. To produce the classification program with the evolutionary algorithm, we need a set of data where we know what the target types are in advanced. This data will be used to create a training file for the evolutionary algorithm. The features are extracted for each chip and entered into a training file along with the correct classification code. The training file is an ASCI file that has the data from a single chip on a single row. Each row has the data listed in the same order to form a set of columns. The first column contains all the classification codes (the result we want from the computer program) and the remaining columns contain the features extracted from the image chip. We usually have thousands of image chips, which means that we have thousands of rows of data in the training file representing hundreds of examples of each of our 10 target types (plus some examples that aren't any of the 10 target types that will get a "not classified" code). This training file is given to e, the evolutionary algorithm and it evolves computer programs that output the correct classification codes (with some errors, of course). The key to making this work is to have meaningful features that can actually help discriminate between the different target types. The evolutionary algorithm determines a good program that can use these features in combination to make correct classifications with low error rates. After testing on an independent test file (where we also know the answers) the resulting programs can be used to automatically classify chips from operational data where we don't know the answer.

It is also possible to adapt the resulting programs in a real time system by running the evolutionary algorithm in the background while the target acquisition process is going on. This can be important when the classification program depends on what terrain type in which the targets are located (e.g. hilly, desert, water, snow, etc.)

G5: What areas of AI are most promising with reference to military applications? G5: How do you foresee the future of AI in the military?

My answer is combined for these two questions. I'm sure that I'm not the best one to answer this question, but I'll give it a try. AI technologies that can aid in automation applications have the biggest potential, it seems to me. At the simplest level this might mean using adaptive algorithms that are capable of modifying themselves within a narrow range to better do an automated task like the target detection problem above. (evolutionary algorithms, neural nets) Expert systems are already having successful applications in equipment maintenance and trouble shooting. I think this application will increase significantly in the future. These systems will help human equipment maintainers make diagnostic and repair decisions. Other expert systems and classification systems will be useful in synthesizing gigantic amounts of operational and intelligence data into understandable packets of information for human commanders. Finally, I think that unmanned weapon systems (air, ground, and sea) will be extensively developed and AI technologies will be gradually introduced to a greater degree as that development matures. These unmanned vehicles will require less direct control from their human operators to accomplish their mission. To accomplish this will require a lot of specialized algorithms for sensing, control, and decision making. In these actual applications some may find it difficult to label a particular approach or algorithm set as falling under the AI umbrella, but the net result will be very autonomous vehicles operating within a set of constraints but not a lot of direct control by their human controllers.

G5: Do you think that humans will ever be taken out of the 'decision loop' during a war? Autonomous fighter aircraft, tanks, even 'generals'? Do you ever foresee a 'Terminator'-type scenario?

At the lowest level of weapons delivery, I do think we will see humans out of the decision loop. Weapons systems will automatically make specific targeting decisions within a set of constraints. These constraints will limit targets to certain locations and types, but the decision to hit this vehicle or that one will be made by the automated system (on an unmanned vehicle or in the weapon itself). I can also foresee automated systems processing intelligence and operational data and even making tactical recommendations, but I expect that humans will be in the loop when the scope of the decision is greater in space (location) and time than the individual targeting decisions I spoke about above. I'm not free-thinking enough to foresee a robotic terminator like the movies.

G5: Do you see any moral or ethical problems with artificial intelligence being used in the military? For example, UAVs that can autonomously designate and destroy targets.

No I don't. We humans rely on all sorts of hardware and software that if they fail will result in undesirable loss of life and property. Automatic systems operate within constraints that are set by humans, but are not perfect. This is true now and will continue to be true as we use AI to expand the capabilities of automated systems. Any moral or ethical issues are more appropriately assigned to the decision to create the mission for the automated system in the first place. It's the mission that has moral and ethical implications, not the technology that implements that mission. A conventional 500 pound bomb can sit harmlessly in a storage facility,or be dropped on a military target in the middle of a battle, or be dropped on a peaceful non-combatant civilian population. The weapon is the same but the mission is entirely different. The same will be true of weapon systems that use AI technology.