Neurax User's Manual

Neurax User's Manual

Neurax is a very powerful horse race outcome predictor which uses the latest in neural network technology combined with fuzzy logic techniques.� Based on its past experience, this artificial intelligence software predicts the most likely outcome of future races.� It has extensive knowledge of all race tracks in North America under varying� track surfaces and weather conditions.�

Seventy five (75) handicapping factors are calculated for each horse.� Then all horses are placed in simulated competition using our neural network (+fuzzy logic) model, resulting in a predicted finish order for each race.

Neural Networks shipped with Neurax

The Neurax product is shipped with multiple neural networks.� Each of these neural networks was trained by our experts using different training strategies.� As a result, each neural network is different from the others with its own particular strong points which are discussed below*.� Some networks are better than others at picking winners, while some excel in their trifecta picks, etc.� It is recommended that users experiment with all the neural networks to determine which offer the greatest benefit at tracks where they bet.�

1.� Prize

This network has been trained to be a general-purpose expert horse handicapper.� Large training sets were used resulting in a very stable network with quite high correct prediction rates for all bet types.�

2.� Leader

This network has been trained to focus on race front runners.� It shows excellent win bet performance.

3.� Exotics

This network delivers excellent trifecta and superfecta bet performance.

4.� Winner

This network delivers excellent win and exacta bet performance.

5.� Hipay

This network has been optimized to pick higher paying winners.� It tends to "gravitate" toward higher paying tickets.�

*Individual network performance was tested on our test set which consists of approx. 1,000 races that were not used during training.

Transparent Unzipping

Neurax operates on two types of data files:� Neurax data files (NRX) which contain information about races that have not yet run (i.e., Daily Racing Form and Track Program information), and Exotic Results data files (XRD) which contain results data for races that have previously run.

NRX and XRD files may be downloaded from Bloodstock Research (BRIS) via Internet or simple communications software.� Usually these files are compressed to save space and lessen transmission times.� Compressed files (ZIP files) are then saved on your computer.�� Uncompressing� (i.e., unzipping) these files will produce the corresponding NRX or XRD file.

Optionally, users may choose to leave these files compressed.� In this case, ZIP files encountered by Neurax will be automatically unzipped as needed.

Neurax treats all ZIP files encountered as equivalent to the corresponding NRX or XRD file that� would result from unzipping.� All ZIP files are displayed as though they have already been unzipped.� When a compressed NRX or XRD file is actually needed by Neurax, unzipping is performed "on the fly" in a manner transparent to the user.�

The decision whether to unzip files after download or to leave them compressed is left up to the user.� Compressed files take approximately 1/4 the disk space of their unzipped equivalents.� Conversely, because compressed files must be unzipped by Neurax before they can be used, performance (of some tasks) will be somewhat slower.

� Files unzipped by Neurax are removed when no longer needed (the original ZIP file is untouched).� The last character (in the file name)� of a ZIP file received from BRIS indicates it's file type.� H indicates a Neurax data file (NRX when unzipped).� X indicates an Exotic Results data file (XRD when unzipped).� To illustrate, Neurax treats file AQU1213.NRX and AQU1213H.ZIP as equivalent (both are displayed as AQU1213.NRX).� Likewise, AQU1213.XRD and AQU1213X.ZIP are treated as equivalent (both are displayed as AQU1213.XRD).

Handicapping Races with Neurax

During handicapping, predicted race results are generated using Neurax's advanced neural network technology.

After downloading the Neurax (NRX) data files from BRIS for the race tracks and dates desired (using communications software of your choice to any folder you like), select Handicap Races... under the Handicap menu.� Simply find the Neurax data files on your computer and add the race cards you want to the handicap list.� Select a neural network to use, and press Handicap.

The Find NRX Files section allows you to navigate through your file system, adding files to be handicapped to the handicap list.� The NRX Files listbox contains a list of all NRX (Neurax data files) in the current folder.� Double-clicking on a folder name goes into that folder.� Pressing the UpOneLevel button goes up one level in the directory structure.� Selecting a different drive letter goes to the root of that drive.� Once an NRX file is located, to add it to the handicap list, simply select it (click on it) and press Add File (alternatively, you may simply double-click on the filename).� Pressing the Search button will search the selected drive, setting the current folder to that containing the most recently downloaded Neurax (NRX) data file�this is useful if the location of recently downloaded files is not known.

The Handicap List section maintains a list of files to be handicapped (i.e., the handicap list).� This list is "sticky" in that all files placed in the list remain in the list until the user removes them (or exits Neurax); this allows easy application of different neural networks to the same set of files.� Selecting a file and pressing the Remove button will remove it from the handicap list (the file itself is untouched).� Pressing the Clear List button will remove all filenames from the handicap list.� If you make a mistake while editing the handicap list, pressing Cancel will undo the most recent changes.� Selecting a file (in the handicap list) and pressing the Edit Scratches button invokes the Scratch Editor which allows editing of late scratches for a given card (the NRX file itself is modified).

One of� several neural networks may be used for handicapping, depending on user preference.� In Single Neural Network mode, all data files in the handicap list will be handicapped using the chosen neural network.��

An Alternate Output Format is available, where all 5 of the standard neural networks (shipped with Neurax) are applied to each card.� Output includes the predicted finish position (rank) and the score using each network in turn.� Horses are listed in program order.

Pressing the Handicap button will handicap all files in the handicap list.

Handicap results are displayed on-screen with one window per track card.� Maneuvering between windows may be done by clicking on the window title bar or by selecting the window at the bottom of the Handicap menu.� At any time, resizing the main window (by maximizing/restoring it or by stretching one of its borders) will cause all handicap result windows to be resized (to fit)�.� The PgUp, PgDown, and arrow keys may be used to aid viewing.�

�By default, the main window is maximized.� To change the default to a standard size window, just restore the main window.� Conversely, to change the default back to maximizing the main window, simply maximize the main window.

Distinct prediction results are generated for DRY track conditions and WET track conditions.*� If the track conditions are DRY at race time, use the DRY track predictions.� If the track conditions are WET at race time, use the WET track predictions.� A WET track is defined to be a track condition that is rated as "Sloppy" or "Muddy".� A DRY track is defined to be a track condition that is rated as "Good", "Firm", or "Fast"�.� (If you must place bets without knowledge of the track condition at race time, it is recommended that you assume DRY track conditions since "Sloppy" or "Muddy" track conditions are relatively rare.)� Keep in mind that the neural network has been trained to take track condition into account�therefore, whenever possible you should bet using the prediction that is appropriate for track conditions at race time.

* Checking the Assume DRY only option, will generate DRY track predictions only.

�"Yielding" or "Soft" tracks are considered DRY.

The Scratch Editor (invoked by selecting a file in the handicap list and pressing the Edit Scratches button) can be used to remove horses from consideration that have scratched�this actually permanently edits the Neurax (NRX) data file itself.� This is the preferred method of dealing with late scratches since a new odds line and finish order will be generated to reflect the new field (during handicapping).

If a horse scratches at the last minute (and you have no computer access), simply cross it off the list (of predicted finish order) and continue as if it were never there; if this method is used, the odds line and score values will be somewhat off since all contenders are used in their calculation�predicted finish order usually will remain mostly unaffected.�

TURF and DIRT races are handled quite differently from one another; consequently, if a race runs on a different surface than designated in the Neurax race data file, it is recommended that the race be skipped.

During the handicapping process, multiple simulations are performed (based on knowledge collected during training) in an attempt to predict this race's most likely outcome.� Each horse is given a score, which reflects that horse's probability of winning that particular race*.� Horses are then ranked by their score values to generate final predicted finish order.

*as calculated by the neural network.� score = win probability * 100.

An odds-line is also calculated (i.e., odds to win for each horse).� A horse's Wodds� value designates the calculated win odds for that horse�horses that run at odds significantly greater than this may signal lucrative betting opportunities.� For example, if a horse is predicted to have a 50% chance of winning it's odds (Wodds) are listed as even money (1:1); if a horse is predicted to have a 25% chance of winning it's odds are listed as 3:1.� In other words, the odds listed are the computed "break even" odds for a given horse.�

�A horse's Wodds value is simply calculated from it's score: Wodds=[1/(score/100)]-1.

The odds line is generated in two stages.� First, fuzzy logic algorithms are used to analyze the (neural network) simulation results producing an initial win probability for each horse.� Second, this initial win probability is processed by performing a non-linear (logistic) regression (via neural network) to give an improved (final) win probability for each horse.�

Neurax data files downloaded close to race time will have track program information in them, eliminating the need to download or purchase a separate track program.� Program numbers, morning line odds, and allowed wagering (e.g., QUINELLA, etc.) are included whenever present.

Glossary of terms always present:

�� PP �� post position

�� Score� �� Horse's overall score� (0-100).� Score values are simply calculated win probabilities * 100.

�� Wodds�� Odds of winning as reported by Neural Network (to $1) [for overlay/underlay detection]

�� Rx �� Today's medication (L=Lasix, B=Bute)

�� Age�� horse's age and sex (f=filly, m=mare, c=colt, h=horse, g=gelding)

�� Wgt �� today's weight carried by this horse

Glossary of terms present if track program information is available:

�� Pgm �� Program #.� Where the program # and the post position are identical, only one number appears.

�� Modds � Morning Line odds (to $1)

Handicap results may be printed to generate hard copy.� While viewing any handicap results window, simply pressing the printer icon (on the main menu bar) will print that card.� To print all cards or groups of cards, select the Print Handicap Results... option under the Handicap menu�users may specify that all results are to be printed or individually select which cards are of interest.� Printer options may be modified via the Printer Setup menu item in the Handicap menu (paper margins are determined programmatically and are not user-selectable).

Neural Network Mapping

Normally a single neural network is applied to an entire race card.� However, advanced users are given the option of specifying which neural network is to be used for a given race type.�

A neural network map (or "NN map") consists of a table which associates neural networks and their corresponding race types.� During handicapping, each race is handicapped using it's corresponding neural network (described in the NN map).�

Race types are designated by specifying a race filter criteria consisting of conditions that must be met.� Conditions may include the track code*, surface (turf/dirt), distance (route/sprint), track condition (wet/dry)� and class (e.g., Stakes) for a race.� If a race� meets all conditions within a criteria, the associated neural network is used.� If a race does not meet the criteria (i.e., one or more conditions are not satisfied), the next criteria is checked (in order) until a matching criteria is found or the NN map is exhausted.� If a race matches none of the criteria in the NN map, it is skipped (i.e., not handicapped).

*Track code where the race was actually run (e.g., the Simulcast track if it was a simulcast) is always used during all filtering operations.

�This allows different networks to be used to generate WET and DRY predictions.

To view/edit NN maps, select Handicap Races... from the Handicap menu and then press the Edit NN Maps button.�� Within each NN map, a comments field is provided as an organization aid�enter a description of the map in any format you wish (it is for display purposes only).� Individual criteria lines may be enabled/disabled individually.� Pressing Reset Map resets a map to default values.� Pressing Cancel discards any changes made (to all NN maps).

DataSets

A dataset is a special database file which is created from an arbitrary collection of Neurax race data files (NRX) and their corresponding exotic results files (XRD).� Once created, datasets can be used to train new neural networks and to generate Profit/Loss analyses to test neural network performance.�

However, before a dataset can be used it must be created.� The dataset creation process simply consists of specifying a folder which contains the NRX and XRD files to be included in the dataset*.� In essence, data files (NRX & XRD) are processed, "producing" a dataset.� Once created, datasets may be renamed and deleted as needed.

*Both the NRX and associated XRD files must reside in the same folder.� Once a dataset is created, the files that were used in it's creation (NRX/XRD) may be deleted, moved, etc. since the dataset is a file separate from those used to create it.

To create a dataset, place the NRX files and their associated XRD files in a folder by themselves and select Create DataSet... under the Database menu.� Locate the folder containing the data files to use, select a name for your dataset* and press Build.� The Append option allows files to be added to the end of an existing dataset.

*Long filenames are used throughout (dataset names and neural net names).

The NRX+XRD Data File Location section allows you to navigate through your file system to locate the folder containing the data files that are to be included in the dataset.� Double-clicking on a folder name goes into that folder.� Pressing the UpOneLevel button goes up one level in the directory structure.� Selecting a different drive letter goes to the root of that drive.

All NRX and XRD files in the current folder are automatically displayed in the Matched Pairs section.� Un-paired files (an NRX file without results or vise-versa) are displayed (in the Left over files section), so you can download the missing files from BRIS if desired.� Un-paired files cannot be included in a dataset, since necessary information is missing.�

Existing datasets may be renamed by selecting the Rename DataSet... menu item.

Existing datasets may be deleted by selecting the Delete DataSet... menu item.�� Select the dataset(s) to be deleted; holding the Ctrl key down will enable multiple selection (dragging can be done to select ranges).� Pressing OK will delete all selected datasets.

Selecting Split DataSet... allows for splitting of an original (source) dataset into a train dataset (used for training neural networks) and a test dataset (used to gauge generalization of trained neural networks).� A specified percentage of the original (source) dataset is copied into the (newly created) test dataset, with the remainder going into the (newly created) train dataset.� The original (source) dataset is left unchanged (i.e., race data is copied).

By selecting Combine DataSets... multiple datasets may be combined (concatenated) to produce a new dataset.� Select the datasets that are to be combined; holding the Ctrl key down will enable multiple selection (dragging can be done to select ranges)*.� (Selecting only a single dataset has the effect of making a duplicate.)� Enter a name for the new dataset.� Pressing OK will begin the process.

*Holding the Ctrl key down and mouse clicking selects/unselects another dataset.� Holding the Ctrl key down, pressing the mouse button (and holding) while dragging will select/unselect ranges.

Two dataset summary reports are available.� An Exotics Report... generates mean and median exotic payouts (for $2 bet) by track;� checking the Copy to Clipboard option, generates a comma delimited version of the report and puts it into the Windows clipboard for easy pasting into spreadsheets, etc.� Payouts vary widely among tracks�it seems prudent not to waste one's time betting exotics at tracks where the likely payoffs are minuscule.�� The DataSet Contents Report... shows dataset contents by card; checking the Copy to Clipboard option copies the text (readable format) report into the Windows clipboard for easy pasting into a word processor, etc.

�

Profit/Loss Analysis

Performance of any neural network may be tested on arbitrary data sets, producing a Profit/Loss analysis for different betting strategies..�

Select Profit/Loss Analysis... from the Database menu� to test a given neural network on a previously created dataset.� You will be asked to choose (from a list) which neural network you wish to use and also (from another list) which dataset you would like to test on.�

Races within a dataset may be filtered by race class�, distance�, track condition, surface, and track initials (BRIS track code)*.� Only races within the dataset that match the filter criteria will be used during testing (i.e., profit/loss analysis).

*Profit/Loss filter criteria are "sticky" in that they remain as set until the Reset button is pressed or the program is exited�this facilitates enumeration.� Routes are races of 7.5 furlongs or longer.

�Selecting "Custom" enables custom criteria setup, where any number of specific criteria may be enabled (only races that match one or more of the checked criteria are included in the profit/loss analysis).

Testing a dataset produces a profit/loss analysis for different bet strategies.� This profit/loss analysis may be viewed on screen and/or printed.� PgUp, PgDown and the Arrow keys may be used to aid on-screen viewing; also, clicking on a given line will display it in blue, acting as a bookmark, to aid viewing.

Enabling the Auto Scratch option instructs Neurax to remove all scratched horses before handicapping each race*.� If this option is disabled, handicapping is performed with all entered horses as described in the NRX file�late scratches are "crossed off " (i.e., removed, after a predicted finish order is generated and win odds are calculated)�.� The resulting finish order is compared against actual race results.

*The NRX and XRD files themselves are left untouched.

�This is analogous to when printed handicap predictions are taken to the track and re-handicapping is infeasible when late scratches occur (due to no computer access).� In this situation, users are instructed to simply "cross off" late scratches (with a pencil) at the track.

If the Force DRY option is enabled, all races are handicapped as though they took place on DRY tracks regardless of actual track conditions.� This is useful if DRY track predictions wish to be used even on WET tracks (e.g., if bets are placed without knowledge or regard of actual track conditions).� If this option is disabled, the race is handicapped using actual track conditions.�

If the Skip TURF->DIRT option is enabled, races which were initially scheduled to run on TURF (as specified in the NRX file) but were later changed to DIRT are skipped (i.e., not counted in the profit/loss report).� If this option is disabled, races which change from TURF to DIRT are counted and are handicapped assuming TURF.

Normally, all odds values indicate actual track odds (i.e., the odds to win at race start).� However if the Use Morning Line Odds option is enabled, all betting strategies relying on odds (for bet placement) use morning line odds instead of actual track odds� (e.g., WIN_2TO1 places a win bet if the morning line odds for the top pick are 2:1 or greater).

�If no morning line odds are available for this horse, the bet is assumed not placed.

�

Enabling the Short Report option generates a profit/loss analysis report without an extended odds or field size breakdown.

Enabling the Copy to Clipboard option places a comma delimited text version of the profit/loss analysis report in the Windows clipboard, allowing for easy pasting into a spreadsheet or word processor.

� Note that some races in a dataset may be skipped, because complete data is not available.

Betting Strategies (used in Profit/Loss Analysis)

This section contains a list of all bet strategies used during a Profit/Loss Analysis and their corresponding meaning.

In the following discussion:� the A horse refers to the predicted winner; the B horse refers to the predicted place horse; the C horse refers to the predicted show horse.

WIN

� Bet A horse (predicted winner) to win.

WIN_SCOREssPLUS

� Bet A horse to win, only if that horse's score value is ss or more.

WIN_OVERLAY

� Bet A horse to win, only if� track win odds for that horse exceed neural network calculated win odds (Wodds value).� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

WIN_OVERLAYn

� Bet A horse to win, only if� track win odds for that horse exceed n times neural network calculated win odds.� i.e., bet is placed only when odds>(n*Wodds).� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

WIN_QFAV

� Bet A horse to win, only if it is a quasi-favorite (i.e., track win odds for that horse are 2:1 or less).� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

WIN_nTO1

� Bet A horse to win, only if track win odds for that horse are n:1 or higher.� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

WIN_FIELDn

� Bet A horse to� win, only if the actual field size (the number of horses at the gate) is n or more.

WIN_ssP_QFAV

� Bet A horse to win, only if that horse's score is ss or more and it is a quasi-favorite (track win odds 2:1 or less).� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

WIN_ssP_nTO1

� Bet A horse to win, only if that horse's score is ss or more and it's track win odds are n:1 or more.� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.�

WIN_ssP_FIELDn

� Bet A horse to win, only if that horse's score is ss or more and the actual field size is n horses or more.

PLACE

� Bet A horse (predicted winner) to place.

PLACE_SCOREssPLUS

� Bet A horse to place, only if that horse's score is ss or more.

PLACE_QFAV

� Bet A horse to place, only if that horse is a quasi-favorite (track win odds 2:1 or less).� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

PLACE_nTO1

� Bet A horse to place, only if that horse's track win odds are n:1 or more.� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

PLACE_FIELDn

� Bet A horse to place, only if the actual field size is n horses or more.

PLACE_ssP_QFAV

� Bet A horse to place, only if that horse's score is ss or more and it is a quasi-favorite (track win odds 2:1 or less).� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

PLACE_ssP_nTO1

� Bet A horse to place, only if that horse's score is ss or more and it's track win odds are n:1 or more.� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

PLACE_ssP_FIELDn

� Bet A horse to place, only if that horse's score is ss or more and the actual field size is n horses or more.

SHOW

� Bet A horse (predicted winner) to show.

SHOW_SCOREssPLUS

� Bet A horse to show, only if that horse's score is ss or more.

SHOW_QFAV

� Bet A horse to show, only if that horse is a quasi-favorite (track win odds 2:1 or less).� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

SHOW_nTO1

� Bet A horse to show, only if that horse's track win odds are n:1 or more.� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

SHOW_FIELDn

� Bet A horse to show, only if the actual field size is n horses or more.

SHOW_ssP_QFAV

� Bet A horse to show, only if that horse's score is ss or more and it is a quasi-favorite (track win odds 2:1 or less).� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

SHOW_ssP_nTO1

� Bet A horse to show, only if that horse's score is ss or more and it's track win odds are n:1 or more.� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

SHOW_ssP_FIELDn

� Bet A horse to show, only if that horse's score is ss or more and the actual field size is n horses or more.

BWIN

� Bet B horse (predicted place horse) to win.

BWIN_OVERLAY

� Bet B horse to win, only if� track win odds for that horse exceed neural network calculated win odds (Wodds value).� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

BWIN_OVERLAYn

� Bet B horse to win, only if� track win odds for that horse exceed n times neural network calculated win odds.� i.e., bet is placed only when odds>(n*Wodds).� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

BPLACE

� Bet B horse to place.

BSHOW

� Bet B horse to show.

CWIN

� Bet C horse (predicted show horse) to win.

CWIN_OVERLAY

� Bet C horse to win, only if� track win odds for that horse exceed neural network calculated win odds (Wodds value).� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

CWIN_OVERLAYn

� Bet C horse to win, only if� track win odds for that horse exceed n times neural network calculated win odds.� i.e., bet is placed only when odds>(n*Wodds).� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

CPLACE

� Bet C horse to place.

CSHOW

� Bet C horse to show.

TOP2WIN

� Bet both the A and B horse to win.

TOP2WIN_SCOREss

� Bet both the A and B horse to win, only if the A horse score is ss or more.

TOP2WIN_nTO1

� Bet both the A and B horse to win, only if the average win odds [(oddsA+oddsB)/2] is n:1 or more.� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

TOP2WIN_FIELDn

� Bet both the A and B horse to win, only if the actual field size is n horses or more.

TOP2PLACE

� Bet both the A and B horse to place.

TOP2SHOW

� Bet both the A and B horse to show.

ACROSS_BOARD

� Bet A horse to win, place and show.

ACROSS_BOARD_QFAV

� Bet A horse to win, place and show, only if it is a quasi-favorite (track win odds 2:1 or less).� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

PLACE_N_SHOW

� Bet A horse to place and show.

PLACE_N_SHOW_QFAV

� Bet A horse to place and show, only if it is a quasi-favorite (track win odds 2:1 or less).� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

B_ACROSS_BOARD

� Bet B horse to win, place and show.

C_ACROSS_BOARD

� Bet C horse to win, place and show.

BC_WIN

� Bet both the B and C horse to win.

BC_PLACE

� Bet both the B and C horse to place.

BC_SHOW

� Bet both the B and C horse to show.

EXACTA

� Straight exacta bet (A horse followed by B horse).

EXACTA_SCOREss

� Straight exacta bet (AB), only if the A horse score is ss or more.

EXACTA_KEYn

� Straight exacta bet, where the A horse is "keyed" with the next n horses.� For this bet to pay, the A horse must win with any one of the next n horses completing the ticket.� {note:� n bets are placed�the A horse is paired with each of the next n horses.}

EXACTA_nTO1

� Straight exacta bet, only if the average win odds [(oddsA+oddsB)/2] is n:1 or more. �If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

EXACTA_FIELDn

� Straight exacta bet, only if the actual field size is n horses or more.

BOX2EXACTA

� Two exacta bets (AB,BA).

BOX2EXACTA_SCOREss

� Two exacta bets (AB,BA), only if the A horse score is ss or more.

BOX2EXACTA_nTO1

� Two exacta bets (AB,BA), only if the average win odds [(oddsA+oddsB)/2] is n:1 or more.� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

BOX2EXACTA_FIELDn

� Two exacta bets (AB,BA), only if the actual field size is n horses or more.

QUINELLA

� Quinella bet (AB).

QUINELLA_SCOREss

� Quinella bet, only if the A horse score is ss or more.

QUINELLA_KEYn

� Quinella bet, where the A horse is "keyed" with the next n horses.� For this bet to pay, the A horse must win (or place) with any one of the next n horses completing the ticket.� {note:� n bets are placed�the A horse is paired with each of the next n horses.}

QUINELLA_nTO1

� Quinella bet, only if the average win odds [(oddsA+oddsB)/2] is n:1 or more.� If the Use Morning Line Odds option is enabled, morning line odds will be used in lieu of actual track odds.

QUINELLA_FIELDn

� Quinella bet, only if the actual field size is n horses or more.

BOX3EXACTA

� Three horse box exacta.

BOX3EXACTA_SCOREss

� Three horse box exacta, only if the A horse score is ss or more.

BOX3EXACTA_FIELDn

� Three horse box exacta, only if the actual field size is n horses or more.

TRIFECTA_KEYn

� Trifecta bet, where the A horse is "keyed" with the next n horses.� For this bet to pay, the A horse must win with any two of the next n horses completing the ticket.� {note:� n!/(n-2)! bets are placed�all combinations of the next n horses filling place and show positions.}

BOX3TRIFECTA

� Three horse box trifecta.

BOX3TRIFECTA_FIELDn

� Three horse box trifecta, only if the actual field size is n horses or more.

BOX4TRIFECTA

� Four horse box trifecta.

BOX4TRIFECTA_FIELDn

� Four horse box trifecta, only if the actual field size is n horses or more.

SUPERFECTA_KEYn

� Superfecta bet, where the A horse is "keyed" with the next n horses.� For this bet to pay, the A horse must win with any three of the next n horses completing the ticket.� {note:� n!/(n-3)! bets are placed�all combinations of the next n horses filling place, show and fourth positions.}

BOX3SUPERFECTA

� Superfecta bet boxing the top 3 horses with the predicted fourth place horse.� For this bet to pay, the top 3 horses must finish 1,2,3 (in any order) and the predicted fourth place horse must finish fourth.

BOX4SUPERFECTA

� Four horse box superfecta.

Training Neural Networks

Users may train additional neural networks using a dataset of their choosing.� To train a new neural network, select the Train Neural Network... option under the Train menu.� Select the dataset you wish to use for training (previously created) and enter a name for your new neural network*.� Pressing Train will begin the training process.

*Long filenames are used throughout (dataset names and neural net names).

Races within a dataset may be filtered by race class, distance, track condition, surface, and track initials (BRIS track code).� Only races within the dataset that match the filter criteria will be used for training.

Training is particularly useful for training neural networks that are sensitive to individual track differences.�

When training a neural network for a particular track, it is recommended that retraining be done every 30 days using a dataset consisting of races and results for the last 30 days at that particular track�in this way, current track conditions may be optimally modelled by the neural network.� The more data that can be accumulated over the last 30 days, the better�therefore, user's are urged to download Neurax data files for the track of interest for all days when data is available, whether or not betting is done on that particular day.

Alternatively, one may choose to simply accumulate Neurax data files from the start of the track season (or as far back as is possible) up to the current date, while retraining using this cumulative dataset every week or so.� Again, the more data files the better.

During training, Neurax starts from a general knowledge base and re-aligns it's neural network to converge onto the given dataset.� In this way, specific information about the dataset is included in the neural network model.� The learning process may be viewed as a continuum progressing from general to specific knowledge.� By increasing the learning pressure, the network is forced more toward dataset specific knowledge (and away from general knowledge).

Alternatively, training may begin from an existing neural network instead of the general knowledge base.� This feature may be useful for "fine tuning" or "updating" a neural network with new data (contained within a dataset).� When applied to an existing network, learning pressure may be viewed as a continuum determining the degree of adaptation to new data that the network will undergo.

For best results, it is strongly recommended that at least 500 hundred or 1,000 races be used for training.� If you use fewer races for training, it is recommended that only the "low" learning pressure setting be used.� With very small amounts of data, "overfitting" can become a problem which results in degraded neural network performance*.�

*Odds Line training requires at least 150-200 races to be mathematically sound; training with fewer races than this is dubious.

If the Favor Winner option is enabled, only horses in finish positions 1 through N (user input) are exposed to the neural network during training.� If this option is disabled, all horses are presented during training.

If the Favor Profit option is enabled, only higher (or lower) paying races are presented to the neural network during training resulting in a network which will tend to "favor" longshots (or favorites).� Only races that have payouts (for a $2 bet) in the range specified are exposed to the neural network during training.� The NOT button may be used to logically reverse the expression.� The AND/OR button may be used to simply toggle the AND/OR status (i.e., change OR to AND and vise versa) without reversing the relational operators.� A blank or zero field value disables that sub-expression (e.g., to select races based solely on trifecta payouts leave the win and exacta fields blank).� Ironically, it seems that enabling the "favor profit" and using only the lower paying horses gives quite impressive overall results; there are many possible combinations, each offering their own benefits and limitations which can be determined by experimentation only.

Since both the Favor Winner and the Favor Profit options reduce the amount of data used during training due to selective filtering, it is recommended that a rather large dataset be used for training when these options are enabled.

After training, it is recommended that a Profit/Loss Analysis be done using the newly trained neural network on the same dataset used for training so that one can get a feel for the effects of training*.� Experimentation is recommended to determine which learning pressure setting gives the best results.� Computer speed is a limiting factor here, since as learning pressure increases one level, the total training time approximately doubles.

*Since the dataset was used for training, results can be expected to be somewhat higher than on novel data.� Ideally, an additional test dataset (not used during training) should be tested (i.e., Profit/Loss Analysis) to gauge how well results will generalize to novel data.

For the really serious horseplayers who wish to train their own custom neural networks "from scratch" (as opposed to "fine tuning" an existing neural network):� starting from the general knowledge base using a "very high" learning pressure with a training set of at least 1,000 races should give the very best custom neural nets.�� Since training on large datasets takes hours, simply letting the software run overnight may be easiest.� If you are enabling the "favor profit" or "favor winner" options, it is recommended that you double or triple the number of races (in the dataset) used for training since these options diminish the total number of patterns available for neural network learning.� In general, using more training data precludes the danger of "overfitting" and increases "generalization" to novel data.� Our experiments indicate that the optimal number of races to use for training lies somewhere within the 1,000 to 5,000 range; exceeding 5,000 races seems to offer little additional benefit.

One may also create neural networks that are optimized for specific bet performance.� First, start with one of the five standard neural networks or start from the general knowledge base.� Then, repeatedly perform training using low (or medium) learning pressure five to ten times (or more if you like), each time using the most recent neural net as the next starting point.� In this way, repeated learning iterations can be achieved.� Finally, use the neural network that maximizes bet performance on an independent test set; delete the other (temporary) neural nets.� For example:� divide a dataset into train and test datasets (via Split DataSet); starting from the general knowledge base, train (using the train dataset) producing net 'A1'; starting from net 'A1' train again producing net 'A2'; starting from net 'A2' train again producing net 'A3', etc; perform profit/loss analysis (using test dataset) on nets 'A1'...AN, picking the one that maximizes bet performance.

We offer only suggestions here.� There are too many possible training strategies to comprehensively list here.� Users are urged to experiment with all available training options to find what best suits their needs.

Track Codes

The following is a list of the most common North American Track Codes:

AKS�� Ak-Sar-Ben�� APX�� Arlington Park

AQU�� Aqueduct�� ATL�� Atlantic City

BEL�� Belmont Park�� BEU�� Beulah Park

BGD�� Blue Grass Downs�� BIR�� Birmingham Race Course

BMF�� Bay Meadows Fair�� BMX�� Bay Meadows

CBY�� Canterbury Park�� CDX�� Churchill Downs

CNL�� Colonial Downs�� CRC�� Calder Race Course

CTX�� Charles Town�� DEL�� Delaware Park

DET�� Detroit Race Course�� DMR�� Del Mar

DUE�� Dueling Grounds�� ELP�� Ellis Park

EMD�� Emerald Downs�� EVD�� Evangeline Downs

FER�� Ferndale�� FEX�� Fort Erie

FGX�� Fair Grounds�� FLX�� Finger Lakes

FMP�� Fairmount Park�� FNO�� Fresno

FPX�� Fairplex�� FSX�� Flagstaff

GGX�� Golden Gate�� GPX�� Gulfstream Park

GSX�� Garden State Park�� HAW�� Hawthorne

HIA�� Hialean Park�� HOL�� Hollywood Park

HOO�� Hoosier Park�� HOU�� Sam Houston Race Park

HST�� Hastings Park�� KEE�� Keeneland

LAD�� Louisiana Downs�� LAX�� Los Alamitos

LRL�� Laurel Race Course�� LSX�� Lone Star

MED�� Meadowlands�� MNR�� Mountaineer Park

MTH�� Monmouth Park�� OPX�� Oaklawn Park

PEN�� Penn National�� PHA�� Philadelphia Park

PIM�� Pimlico�� PLA�� Playfair

PLN�� Pleasanton�� PMX�� Portland Meadows

PRE�� Prescott Downs�� PRM�� Prairie Meadows

RDX�� River Downs�� RET�� Retama Race Park

RKM�� Rockingham Park�� RPX�� Remington Park

SAC�� Sacramento�� SAR�� Saratoga

SAX�� Santa Anita�� SOL�� Solano

SPT�� Sportsman's Park�� SRX�� Santa Rosa

STK�� Stockton�� SUF�� Suffolk Downs

TAM�� Tampa Bay Downs�� TDN�� Thistledown

TIM�� Timonium�� TPX�� Turfway Park

TRM�� Trinity Meadows�� TUP�� Turf Paradise

WDS�� Woodlands�� WOX�� Woodbine

YMX�� Yakima Meadows

Copyright Information:

Neurax may not be copied in any form, except for personal backup purposes.� Neurax's race predictions (handicap results) may not be sold or publicly distributed in any form.�

Disclaimer:

Past performance is not a guarantee of future success.� Individual results may vary.� The user assumes the responsibility of verifying all mathematical assumptions used.� Great Inspirations Software is in no way responsible for any losses incurred by the use of its Neurax product.

Acknowledgements:

Neurax uses special compression and de-compression utilities from the Info-ZIP group (zip.exe and unzip.exe) to process ZIP files.� Info-ZIP's software (Zip, UnZip and related utilities) is free and can be obtained as source code or executables from Internet/WWW sites, including http://www.info-zip.org/pub/infozip/; see infoziplicense.txt (shipped with this product) or ftp://ftp.info-zip.org/pub/infozip/license.html for Info-ZIP�s licensing agreement.