Intrinsically motivated Reinforcement Learning
This is my research in Machine Learning. I worked under Prof. Satinder Singh and Prof. Andrew G. Barto in this research during Fall2003, Winter2004 and at the beginning of Summer 2004. We published two papers in ICDL2004 and NIPS2005. I implemented the algorithm in Java. The domain in the screen shots is the Playroom Domain. The playroom have a number of objects: a light switch, a ball, a bell, two movable blocks that are also buttons for tuning music on and off, and a toy monkey that can make noise. The agent can sense what objects are under its eye, hand and marker. The agent has the following actions available, 1) move eye to hand, 2) move eye to marker, 3) move eye one step to north, east, west or south, 4) move eye to a random object, 5) move hand to eye, 6) move marker to eye, 7) move eye to hand, 8) move eye to marker. In addition, if both hand and eye is at the ball, the agent can throw the ball to the marker. If both hand and eye is at the light switch, it can flick the light switch. Objects in this domain can interact with the agent in various ways, such as when the agent throws the ball to the bell, it makes sound, when the agent flick the light switch, the light turn on/off, etc. Please take a look at the NIPS paper for more detailed about the domain and the algorithm.
