Acknowledgments: This project was funded by the Singapore-MIT GAMBIT Game Lab.
The repository includes codes and compiled executables to create simple collaborative games.
- CAPIRSolver: C++ codes used to compute behavior policies.
- Unity codes: The game codes
- Compiled game executables: Mac OS X and Win32
- CAPIR User Manual
The following demo shows how to
- create a game level using Tiled,
- use the codes included to obtain the behaviors for the AI, then
- run the game to play-test the level just created
The following demo demonstrates the intelligence of the AI
We address the Lead-Assistant Collaboration problem that characterizes an autonomous assistant aiding an agent in achieving multiple goals in a known environment; the main source of uncertainty is the agent's drifting intention and behavior model in achieving the goals. We propose the CAPIR framework (Collaborative Action Planner using Intention Recognition) to solve this problem by tracking the assisted agent's intention using inverse planning (a concept inherited from cognitive science), and constructing actions based on the assistant's belief on the agent's intended goal and its knowledge of solving each goal (modeled as a Markov decision process). For complicated goals that yield large planning space, simulation-based algorithms are devised to approximately solve the goals online. Evaluation with human players in an experiment game shows that the assistant's resultant course of actions is near human-level.
- Truong-Huy D. Nguyen, Tomi Silander, Lee Wee Sun and Leong Tze Yun (2014).
Bootstrapping Simulation-based Algorithms with a Suboptimal Policy.
ICAPS 2014. PDF - Truong-Huy D. Nguyen, Lee Wee Sun and Leong Tze Yun (2012).
Bootstrapping Monte Carlo Tree Search with an Imperfect Heuristic.
ECML/PKDD 2012. PDF - Truong-Huy D. Nguyen, David Hsu, Lee Wee Sun, Leong Tze Yun, Leslie Pack Kaelbling,
Tomas Lozano-Perez, Andrew Haydn Grant (2011).
CAPIR: Collaborative Action Planning with Intention Recognition.
AIIDE 2011. PDF