Counterfactual Simulation and Omissive Causation
Abstract
In his discussion on the frame problem in artificial intelligence~the challenge of determining what an agent needs to represent as r"emaining unchanged after an action or event takes place~philosopher Daniel Dennett introduced the imaginary example of R1, a robot w""hose only task is to fend for itself. One day, designers arranged for R1 to learn that its battery, its only source of energy, was l"ocked in a room alongside a bomb set up to explode soon. There was also a wagon inside the room. R1 s program allowed it to generate the hypothesis that placing the battery in the wagon and pulling out the wagon from the room would allow the battery to be removed" from danger. Unfortunately, the bomb was also on the wagon, and when R1 removed the battery, the bomb came along with it, exploded"" and destroyed the robot.In this dramatic illustration of the frame problem, R1~s behavior exemplifies the more specific difficult"y of modeling successful causal explanations and predictions in autonomous agents. For even though R1 successfully predicted the eff"ect that pulling out the wagon would have had on the battery, it neglected to take into account the effect of failing to remove the"" bomb first. At first sight, both kinds of causal relationships~doing something and failing to do something~appear very similar. How""ever, the last few decades of research in causation and causal modeling have demonstrated that the most successful accounts of causa""tion have a hard time accommodating what is known as omissive causation, a kind of causal relation in which the failure of an event" to occur causes a certain effect.A promising strategy to tackle the difficulties in understanding omissive causation consists in" exploring the neural and cognitive mechanisms involved in thinking about omissive causes. Unfortunately, very little is known about" the psychology and neuroscience of representing and reasoning about omissions as causes. The current proposal seeks to contribute t"o filling this void in the literature, and to expand our knowledge about the cognitive psychology, neuroscience and computational mo""deling of omissive causal reasoning.The current proposal tackles this general objective via three specific aims. First, we seek to" conduct a series of behavioral and neuroimaging studies employing both ERP and fMRI to determine whether people are more likely to" think of omissive causation in terms of a process (e.g., as involving energy transmission) or in terms of counterfactual dependency"". Second, we seek to conduct a number behavioral and neuroimaging studies to fully explore initial evidence suggesting that norms~mo""ral, conventional and statistical~bias people~s decisions when choosing possible counterfactual events as being the relevant would-b""e causes in cases of omissive causation. Finally, in collaboration with the team at NRL, we seek to incorporate the results of our e"mpirical investigation to refine promising computational models of human counterfactual and causal reasoning. The empirical and computational results that will emerge from this project promise to help to improve extant models of causal prediction and explanation in autonomous agents.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 09, 2017
- Source ID
- N000141712603
Entities
People
- Felipe De Brigard
Organizations
- Duke University
- Office of Naval Research
- United States Navy