Robust Collaborative AI to Enable Effective Human Decision Making under Uncertainty and Time-Stress
Abstract
Despite promising advances in AI assistants, developing intelligent agents capable of collaboratively planning with human decision makers remains a challenging problem. We propose to develop a collaborative artificial intelligence (CAI) that is capable of assisting a human decision maker during planning and execution of a mission in dynamically changing and uncertain environments, using disaster response in maritime scenarios as a notional evaluation domain. To address the multiple technical challenges of the project, wewill further develop and extend a novel neuro-symbolic approach that combines neural network based learning with reasoning over human-interpretable decision trees. The proposed CAI is designed to provide bidirectional, mixed-initiative decision support through a dialog-based human-agent interface. The CAI will provide support both in the planning and execution phases. In particular, via interaction with a decision maker the CAI will a) suggest a mission plan given the current state of the environment and mission parameters and requirements; b) critique a current plan and provide suggestions for improved courses of action or alternative plans; c) provide responses to requested information about the state of the environment, available resources, expected task outcomes, and related information; and d) provide advice to repair existing plans due to environmental perturbations, loss of resources, or unexpected task outcomes. The CAI will maintain a library of human-understandable sub-plans to achieve mission sub-goals. Individual sub-plans will be reused in different missions or scenarios to support generalization of the CAI. The CAI will learn sub-plan policies through reinforcement learning. Learned sub-plans will be converted to human-interpretable decision trees using a verifiable policy extraction algorithm. The CAI will learn to both construct mission plans via combination of appropriate sub-plans and to model the consequences of sub-plan execution through reinforcement learning; these learned elements will be leveraged by the CAI to perform its various capabilities during planning and mission execution. We will develop a user interface and natural language dialog system to support interaction between the CAI and human decision maker. The user interface will infer the intents of the human during queries or other interactions, and will perform one or more of its supported capabilities to generate responses. The natural language dialog system will enable the CAI to generate responses to user queries in natural language. Additionally, the CAI will initiate interactions while monitoring plan development and execution, based on observed environmental perturbations or expected undesired outcomes of acurrent plan. Such interactions include suggested plan improvements, the resolution of conflicts the CAI may encounter, or the selection of quantitatively equivalent plan choices. A corpus of domain-specific dialogues will be collected to train the dialog system; training will be augmented with existing, publicly available corpora used in natural language understanding. The CAI will be trained and evaluated on emergency response scenarios using two simulation engines (Ocean Simulator and CHAMP Shoreline Simulator), to demonstrate CAI decision support capabilities, evaluate performance of individual CAI components and human-CAI teams, and to demonstrate the ability of CAI to generalize to different mission instances, scenarios, environments, and simulators.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 11, 2023
- Source ID
- N000142312840
Entities
People
- Katia Sycara
Organizations
- Carnegie Mellon University
- Office of Naval Research
- United States Navy