Project on Physical Training

Introduction

• What is the desired goal of the expert system; how will appropriate conclusions be expressed?
• What variables are relevant to the application?
• How will we represent data?
• Are there some intermediate variables that would be helpful?
• How will knowledge be acquired from an expert? How will the expert provide feedback concerning the validity of rules and/or results?
• Can we identify a database for checking the results of an expert system?
• Within the expert system itself, what type of inference is appropriate (e.g., forward or backward chaining), what type of user interface might be appropriate, and what programming language or software tools are available?

• What specific inputs are relevant to the application? Will those inputs be discrete or continuous?
• What output(s) are desired? Will each output have continuous or discrete values?
• What interior nodes are appropriate?
• What connections should be present between nodes?
• How should weights be determined or learned?
• Can we identify a database for machine learning and for checking the results of an expert system?

This project focuses on a real-work application: the development of an expert system and a neural network to help train middle-distance runners. If the resulting projects are sufficiently successful, they may be used by the Coach Freeman and the Grinnell coaching staff. At the very least, the projects should provide insights into this application domain, so that more extensive AI programs may be developed in the future.

Altogether the training of athletes is multi-faceted, including the development of long-term training models and goals. These long-term goals guide the development of short-term workout schedules, which include specific day-by-day training plans for runners. Ideally, daily workouts should allow athletes to progress along an optimal, progressive curve toward long-term goals.

Overall, running ability and skill may be categorized by major levels, and an athlete would like to progress steadily from one major level to the next. Also, since major levels of achievement include various sublevels, the ideal training program would allow athletes to progress through sublevels appropriately.

In an academic setting, however, the load or intensity for a specific workout must be subject to several variables, such as:

• academic stress, and
• amount of sleep.

• an athlete's anaerobic threshold,
• heart rate at rest, and
• volume of oxygen required for a workout.

• intensity of the workout (e.g., speed, percent of maximum effort, length of recovery periods),
• volume of the workout (e.g., miles run per week, length of individual activities), and
• density of the workout (e.g., number of sessions in a given period of time -- sessions/day or sessions per week).

• Increase training goals by 3 or more sublevels,
• Increase training goals by 2 sublevels,
• Increase training goals by 1 sublevel,
• Do not change the training schedule,
• Decrease training goals by 1 sublevel,
• Decrease training goals by 2 sublevels,
• Decrease training goals by 3 or more sublevels,

Some Details

As previously noted, in order to produce something of possible value within the time constraints of the course, the project will focus only on adjustments to an existing training schedule. Thus, the project will assume that a coach already has determined an appropriate general workout schedule, which will allow athletes to meet their long-term goals over a season. The determination of such an overall schedule is beyond the scope of what can be done in this class

As with the development of many expert systems, a major difficulty here is the determination of what factors should have what effects. This is the domain of experts in the field -- in this case, Coach Freeman and the athletes. To identify such information, the class will develop a questionnaire, asking for known outcomes for various values of variables (as noted below). Specifically, each athlete will be asked to provide information on current experiences and on any patterns that seem common in the athlete's experience. Similarly, Coach Freeman will be asked to record common patterns which he has observed.

With this information, the class's challenge will be to formulate rules which may predict the outcomes from the variables.

To help formulate rules, the survey results will be stored in a standardized format. TMYCIN will be modified to read each case, apply the rules, and print both the rule-based result and the expert's result. Thus, after possible rules have been formulated, they can be tested against the known cases to determine how well they yield desired results. Such automated testing can greatly facilitate the rule-development process.

Additional details are covered under the "specifications" section that follows.

Variables:

Variable	Abbr.	Values
Normal Miles	Nor-Mi	Unknown	< any integer >
Week Miles	Week-Mi	Unknown	< any integer >
Hard Days	H-Days	Unknown	< any integer >
Volume	Volume	Unknown	very-light	light	normal	heavy	very-heavy
Intensity	Intens	Unknown	very-light	light	normal	heavy	very-heavy
Density	Density	Unknown	very-few	few	normal	many	very-many
Sleep/Rest	Rest	Unknown	none	very-little	little	average	excess
Recovery	Recovery	Unknown	much-shorter	shorter	normal	longer	much-longer
Acad. Stress	Stress	Unknown	very-light	light	normal	heavy	very-heavy
Motivation	Motiv	Unknown	weak	average	strong
Heart Rate	HR	Unknown	down-3%	normal	up-3%	up-6%	up-10%
Last Race	l-race	Unknown	poor	below-avg	average	good	very-good	pers-best
Adjustment	Adjust	Unknown	far-below	below	normal	above	far-above

Data File:

• The results for one case will be on a single line.
• The overall format for one case will be a list:

  (identifier  (association list) adjustment)
  

• The identifier for a case will have the format: casenn where nn is a sequence number.
• The association list will consist of (variable abbreviation, value) pairs, such as

  ((volume light) (intens average) (density few) (HR normal))
  

  Variables where the values are unknown will not be included in the association list. Values will follow those prescribed in the above table, except that the percent signs for heart rate will not be recorded.
• A sample case entry might be:

  (case23 ((volume light) (intens light) (density few) (rest little)) below)
  

Results File:

<234567890123456789012345678901234567890123 -- reference numbers>
Case Identifier     Predicted Adjust.   Expert Adjust.

The first line of the output file will include titles of the columns. The rest of the file will contain one case per line.

Note: the expert conclusions were reviewed on Sunday, April 26. Corrections reflect errors identified at that time.

To use the current data files, follow these steps:

1. Copy the default rules file ~walker/public_html/courses/261/track/template-alt.rev
   to your account.

2. Experiment altering rules by adding, deleting, or modifying rules in your copy of template-alt.rev.

3. Run tmycin with your rules by:

   • Starting LISP with the acl command.
   • Loading your rule file with the statement (load "template-alt.rev").
     (Note that tmycin should be loaded automatically.)
   • Running the program by typing (placement) or (adjustment)
   • Using option i for individual placement experiments, or
     using option f to read and write to default files
     the default input file will be ~walker/public_html/courses/261/track/survey-data
     (file survey-data.rev duplicates this data file).
     the default output file will be tmycin.output in your local directory.
   • Leaving the menu with the q option.

4. To run the file-output analyzer, follow these steps:
   • Run LISP as above.
   • If needed, load the analyzer:
     (load "~walker/public_html/courses/261/track/analyze.lsp")
     (This also is loaded automatically by template-alt.rev.)
   • Run the program by typing (analyze).

5. Run the analyzer with the command (analyze)

6. You will need to delete or rename your old tmycin.output file before you can run the program again.