Feedback 16-07-2012

Progress

• Implemented a continuous state space variation of the light-dark domain:
• State space is a subspace [minX, maxX], [minY, maxY] of R^2
• Goal is a square of size 1
• Actions: discrete, with some additive Gaussian noise
• Wrap-around effect: if agent leaves on one side, it enters on the opposite side
• Observations: same as in discrete state space variation
• Implemented a belief approximation:
• Bivariate Gaussian with three parameters: meanX, meanY, variance
• Updates are done with a Kalman filter (simplified to the specific Gaussian and the environment)
• Matrix computations are performed with apache commons math
• Bivariate Gaussian is realized with a subset of classes from jahmm
• Implemented a transposition tree for this Gaussian:
• Instead of maintaining and updating a 1D belief during a simulation, the algorithm maintains and updates the three parameters of the Gaussian
• Searches for split points along the Gaussian's three dimensions
• Selects the best F-test among all dimensions for a split