A Neural Net For Terrain Classification Jackie Soenneker
Overview of SOM Self-Organizing Map Neural Net Has a grid of neurons Each neuron has a weight vector For each input vector there is a “winning” neuron The winning neuron and its neighbors are adjusted to better match the input 1 2 3 7 8 9
Dimensions & Distance Functions Dimension: how many neurons to use Default is 4x6 (24 neurons) I’m using twice as many neurons as terrain classes Distance Function: how far apart are 2 neurons? Link Distance (default) – number of links between the neurons Euclidean Distance – straight-line distance between the neurons Manhattan Distance – “follow the grid” distance between the neurons’ vectors
Topologies Topology: how are the neurons connected? Topology doesn’t seem to effect learning very much Hextop is the default and the one I’m using Hextop Gridtop Randtop
Learning Phases SOM learning has two phases Ordering Phase (first phase) large learning rate quickly fits the neurons to the general distribution of the input space There are 2 Ordering Phase parameters Learning rate – 0.9 (default) Number of steps – 1,000 (default); 2,000 works better The number of OP steps should probably grow proportionally to the number of neurons
Learning Phases con. Tuning Phase (second phase) small learning rate fine-tunes the neurons to fit the input space more precisely There is one Tuning Phase parameter Learning rate – 0.02 (default)
Summary Dimension: twice as many neurons as terrain classes Distance Function: Link Distance (default) Topology: Hextop (default) OP Learning Rate: 0.9 (default) OP Steps: 2,000 (probably increase w/ more nodes) TP Learning Rate: 0.02 (default)