CT: trying it with humans

The Colored Trails platform was designed to provide ease and flexibility in developing and deploying human subject experiments for the study of decision-making behavior. It has hitherto been used to explore a diverse set of questions including, but not limited to, the following:

• how humans negotiate in sharing and using resources,
• how people model their opponents,
• whether and how reciprocity affects behavior,
• the interplay between reducing uncertainty and postponing gains,
• studying the reasoning patterns of agents in games,
• etc.

Experiments are the only way

Such experiments serve to illuminate and verify (or discard) various theories of decision-making. Over the last 15-20 years there has been a growing literature in behavioral and experimental economics (as well as the younger field of neuroeconomics) that attempts to quantify how closely actual humans behave to formal theories of choice. Attempting a crude summarization of its results, this line of work has shown that human behavior is affected both by utility maximization, as well as emotional, perceptual and involuntary/unconscious factors. Hence, it becomes even more important to test any theory by use of experiments to guide newer research towards better and more useful models.

What kind of experiments?

If experiments "are the way," the question arises: Are all experiments equally good?

The answer is: No.

We believe that, if our models of human behavior are going to be useful, they need to be tested in situations that capture the complexity and richness of day-to-day, real-world phenomena. If the domain is too simple, behavior tends to become artificial and classical theories of choice are more easily confirmed, which gives an inaccurate perception of their applicability. On the other hand, if too much complexity is introduced, it becomes excessively difficult to isolate factors relevant to subjects' decision-making and quantify their effect. The solution is to go for something "in the middle."

Colored Trails accomplishes this by having humans interact through a board game that is sufficiently simple to study, but adequately complex to capture important effects. The game features subtle elements like resources of various types, players with private information and beliefs, etc. As a matter of fact, the researcher/developer is completely free to customize the game, increasing or decreasing its complexity according to his/her needs.

Nuts and bolts

CT is a client-server platform: the server maintains a list of "games" currently underway. For each game, a list of connected clients, the board configuration, the resources (chips) held by each client and any other private information is maintained. The server is also responsible for sending and receiving communication messages, advancing the game and maintaining the clients informed of any changes.

Clients can either be "human" or "computational." Human clients are basically GUI applications in which the board is displayed and any moves available to the human decision-maker are presented through an appropriate interface. Computational clients are software agents that make the same decisions like humans using an algorithm (e.g., a learning technique, a heuristic or an optimization technique)---no UI is involved.

Finally, CT also has a controller, which typically resides with the server. The controller is responsible for executing an experiment. It may load up pre-recorded game settings, pair (and re-pair) connected clients together, initiate and terminate games, etc.