In this course, we tried to broaden your horizons for thinking about computer programs. Most of you have been trained (very well) in how to apply quantitative techniques such as machine learning, neural networks, and statistics to solve very sophisticated problems.
We have shown that qualitative factors can be incorporated in your programs to complement your quantitative analyses. These topics include:
Behavioral economics, such as risk aversion, framing, and anchoring.
Qualitative arithmetic, to develop subjective interpretations of numbers as good, bad, better, or worse.
Emotions, that reflect the state of goal pursuit and communicate better with the user.
Explanations as a key component of a decision making program.
Guest speakers who provide real-world perspectives on a range of issues.
Spectrum of programming requirements. The homework assignments required you to explore Monte Carlo simulation (hw1), financial analysis explanations (hw2), and computer technology recommendations (hw3). These assignments were designed to give you increased autonomy to allow you more freedom. The final project has the fewest constraints - you get to select the domain and problem.
Cognitive Modelling. The qualitative approach is based on trying to emulate human behavior. We have encouraged you to think about artificial intelligence from a psychological perspective that can sometimes help solve or identify problems that do not readily admit to quantitative solutions.
Machine learning. I assume that you have been exposed to ML through other courses. I wanted to focus on topics not included in those other courses.
Natural Language Generation. Below I note that language generation is a key part of VOTE. I had planned on a guest speaker, Kris Hammond, to discuss this. Unfortunately, Kris had a last minute conflict and could not speak. However, he has a TED talk on language generation and another superb TED talk on artificial intelligence. In 20 minutes, he says most of what I tried to say all semester.
The GBMD book concludes with its own summary of topics.
We have presented a general model of goals and interpersonal relations in a specific task context of decision making. The major points of this work can be summarized as follows.
Decision Trade-offs. Typical AI planning programs assume that the agent must achieve all of its goals. We claim that this assumption is unrealistic. Agents have a multitude of goals and limited resources. Agents must make trade-offs and compromises that may result in the abandonment of some goals.
Goal Decomposition. Typical AI planning programs treat goals as unitary wholes, that index plans and cases that relate to the achievement of goals. We assert that goals may be decomposed into more primitive elements that provide a basis for comparing and ranking goals. Reasoning about goals is more involved than simply devising plans for achieving goals. Agents must reason about which goals to pursue in the first place.
Principle of Importance. The importance of a goal is proportional to the resources that the agent is willing to expend in pursuit of that goal. The principle of importance applies to the complementary tasks of planning and understanding. For planning, an agent allocates resources according to the relative importance of goals. In understanding, we infer importance based on how other agents allocate resources.
Resource Decomposition. Just as there are a multitude of disparate goals, we claim that there are a great many different types of resources, such as time, money, and physical objects. These resources can be decomposed into primitive elements that provide a basis for comparison and reasoning.
Cognitive Resources. In addition to the more tangible resources, there are cognitive resources, including attention and memory, which can be allocated in the pursuit of goals.
Cooperating Agents. Typical AI programs assume that agents act independently of other agents. We claim that it is more realistic to assume that agents will interact with other agents through interpersonal relationships.
Principle of Interpersonal Goals. Adopted goals are processed uniformly as individual goals, with a priority determined by the importance of the relationship. This principle provides a model for a number of interpersonal phenomena including counterplanning, secondary relationships, and persuasion.
Organizational Relationships. The principle of interpersonal goals may be extended to include goals adopted from institutions and organizations.
Goal-Based Decision Making. We suggest an alternative to quantitative decision analysis by proposing a model of decision making based on an agent’s goals.
The preceding principles served as the basis for the VOTE program which simulates congressional roll call voting. VOTE has the following distinctive features.
Real World Task Domain. VOTE uses data about real bills, issues, groups, and members of Congress to simulate actual roll call votes.
Multiple Decision and Explanation Strategies. VOTE uses a number of modular decision strategies with associated explanation strategies. An indirect strategy serves to increase the level of analysis for all the other strategies.
Natural Language Generation. VOTE provides robust natural language generation both for explaining decisions as well as for interpreting the contents of the database.
Robust Software Infrastructure. VOTE includes an object oriented database system that facilitates the access and update of the various knowledge bases. VOTE’S database system provides a range of features including interactive editing, help, and retrieval, data dependencies, undo, spelling correction, and compilation.
VOTE’s performance in this domain demonstrates the computational feasibility of our model of goal-based decision making.