So, I’ve been hard at work, sometimes so much so that I have neglected to update and record my progress.

Today I spent some time trying to discover the capabilities of the ER1’s speech recognition system. This proved to be quite a bit more difficult than I originally thought. First of all, the speech recognition system which is built into the Robot Control Center (RCC) requires the use of a microphone (obviously) which I did not have on hand. Once I located an adequate microphone and had it configured and adjusted to function with the Windows XP ER1 laptop I realized that the RCC speech recognition function is an extension of the operating system’s speech recognition engine (which was not installed on the ER1 laptop). Thankfully by locating an ancient Microsoft Office XP install disk I was able to install and setup the speech recognition engine and do some of the training examples to improve the recognition. When the system was finally running and functional in the OS, I was able to run the RCC speech recognition behavior with good results. I had the ER1 listen for me to say the phrase “Good Morning” at which point it responded by saying “I just heard you say ‘Good Morning'”. It was pretty satisfying to have the robot hear my voice and respond, I must say. However, there seems to be no functionality (at least no documented functionality) to allow the speech recognition system in the RCC to be controlled or interface through the telnet interface which I am currently using to autonomously control the ER1. This speech recognition problem may be reviewed in the future.

Now to the exciting stuff. Working on the machine learning algorithms. I spent the last couple of days working to write some general decision tree building functions in Matlab, with some success, but lots of frustration. In general a decision tree is a type of regression that uses statistical weights for different inputs to choose a final output class. Troubleshooting guides are a basic type of decision tree that attempt to locate the source of a technical problem. In my system I am attempting to take a labeled collection of inputs and outputs and build a decision tree in Processing that can be implemented on an Arduino. The heavy lifting of calculating all the statistical weights and information gains  make sense to be implemented in Processing because a PC is much more powerful than an Arduino. Then the Arduino (which can be part of a robot, electronic art installation or anything else someone may create) can benefit from the use of the decision tree to make it smarter (I hope).

The design so far is as follows: Input (arbitrary number of attributes (like temperature or humidity) which can each take on some unique arbitrary number of discrete levels (like hot, cool, cold) with all levels represented by integers 1 to n, where n is the number of levels that attribute has. Output (arbitrary number of output states/classes (like yes/no or left/right/center) with all output states also represented by integers 1 to n, where n is the number of output classifications. The data used to train/build the tree must be some collection of ordered sets of attributes and output classes, where a greater number of training data has the potential to build a more accurate tree. A couple of relatively simple equations are used to calculate the values needed to build the tree.

The entropy represents how much variability/randomness a particular attribute contains. An entropy value of 1 means that the attribute has no effect on the output. An entropy value of 0 is deterministic, and leads to a single decision (output classification) and thereby determines when a branch of the tree will end in a terminal leaf. The gain value determines the information gain a given attribute provides based on the current branch it is on. The gain values are used to determine the branch nodes and hierarchy of the tree. The “i” above represents the possible output states and the “v” represents the possible values an attribute can take on.

So far in Processing I have an architecture for the tree and functions to calculate entropy and gain, but the difficult task of making the entire thing recursive and adaptable to arbitrary data sets, and producing some output that is usable by the Arduino is still to be done.