There is a somewhat hacky solution, but it requires some additional code not part of NuPIC. It should be possible, by plopping a standard deep network on top of HTM. So you can use the HTM to direct which sub-networks of the deep network you train (by turning off sections of the network that go through "0" bits). You can then use this to represent the value function. You can also then backpropagate high values to the inputs to get actions.
(this is what I am working on at the moment as my research project). Alternatively, you can just wait for the real sensorimotor integration ;) On Mon, Jan 5, 2015 at 1:33 PM, Matthew Taylor <[email protected]> wrote: > NuPIC currently does not implement behavior. Work is ongoing on > sensorimotor integration into NuPIC. I don't think NuPIC can be > applied to self-driving car problems without this being completed. > --------- > Matt Taylor > OS Community Flag-Bearer > Numenta > > > On Sat, Jan 3, 2015 at 5:52 AM, Dinesh Deshmukh <[email protected]> > wrote: > > Hi > > > > I dont know whether to discuss this here,if not then i am sorry guys. > > > > I would like to know if anyone have any ideas of implementing HTM > algorithms > > in self driving car simulation in any way.The simulation complexity is > very > > simple. > > > > If it is complicated then please explain why.Thank you. > > > > > >
