search

calumroy / HTM

HTM
3 stars 0 forks source link

command Input #4

Open calumroy opened 10 years ago

calumroy commented 10 years ago

Currently the input to the command layer is from the layer below. This won't work as only the next input will reinforce the current command. The command layer needs to reinforces it's commands with a different input then the layer below.

calumroy commented 10 years ago

If the command layer (layer 5) receives inputs from higher regions as well as the new inputs this may work.

Command and sensory connections between layers (work in progress)

https://www.draw.io/#G0B1Pww-SuoSGgU3VyX2xhN3RXQUk

calumroy commented 9 years ago

Another idea that will allow this to work is randomly assigning cells to different output actions. This has the advantage that when a column is activated by an input then all the cells in the column could potentially become active. The cells that do become active are the ones that where in a predictive state. If there is enough cells in a column then one column could potentially output a large range of actions. The action chosen is the action associated with the cell that was predicting and therefore gets activated. The total output should just be the culmination of all the active cells associated actions.

This gives the desirable ability for a HTM region to have any input and still output any action. For example if an input representing a T junction where a robot has driven to, then the output can still be "go left" or "go right" or any other action the robot might want to take. The action selected by the region will depend on which cells where predicting. The cells that are predicting can be influenced by the new input, the feedback and the contextual information from previous inputs.

calumroy commented 9 years ago

Related to Qlearning thalamus implementationissue #8.