Casting And Role Management Equality System is a software package which assists in the casting of roles in theatrical productions. It has been released open source under the GPLv3 license, with the exception of the research paper Research.pdf which is copyright David Lang (2021).
The research paper is included in this repository as the CARMEN software package was the deliverable of this research project, and may be freely read for interest and further research purposes as long as proper attribution is included in any further published works.
Abstract from Research.pdf
Community and not-for-profit theatre is a common hobby for all ages, but casting people into roles appropriate for their ability level can be a challenging task, particularly when considering large groups and alternating casts of children.
This project sets out to evaluate the use of SAT solving and neural network based algorithms for the selection of cast and allocation of roles in theatrical productions, and incorporate the successful concepts into a working software solution.
In order to implement the required algorithms, the boolean satisfiability problem (SAT) and artificial neural networks (ANNs) have been researched. A critical review has also been completed of the software packages previously used by Cumberland Gang Show (CGS), a production company which is used as a case study.
The hypotheses of this research are that SAT solving algorithms will improve the balance of talent between alternative casts, and that machine learning neural network models will improve the accuracy of recommendations and spread of roles allocated to applicants, over the previously used heuristic algorithms.
Six SAT based algorithms for cast selection have been implemented and assessed, with five of them found to improve on the heuristic baseline, and four improving on the historical cast selection as performed manually by the production team. This proves that SAT solving algorithms will improve the balance of talent between alternative casts.
Three ANN based algorithms for role allocation have been implemented and assessed, with various parameters totalling 32 neural models. All of these were found to improve accuracy over the heuristic baseline, proving that neural networks will improve the accuracy of recommendations, however the spread of roles was found to be less, disproving that hypothesis. A thorough discussion is included on the reasons why this may have occurred, and future research which could be conducted in this area.
The completed software package, implementing the successful algorithms, has been named “CARMEN: Casting And Role Management Equality Network” and released open source, under the GPLv3 license. The user acceptance testing by five members of the CGS production team found it was faster, easier to use and more intelligent than the previous software, with all five recommending it for use by other production companies.
This deliverable, and the research that went into it, has improved the quality of the casting process of theatrical productions. It will first be used for a real world production by the Cumberland Gang Show in December 2021, as they start auditions for their July 2022 performance, and will help the volunteer production team to share roles fairly amongst applicants, leaving them to focus on what they do best; developing the talent, creativity and passion for the arts of the youth (and youth-at-heart) of Sydney, Australia.
- Find the latest release and download setup.exe AND Carmen.msi
- Run setup.exe, and follow the prompts to install any pre-requisites
- Once you reach "Welcome to CARMEN setup", press "Next"
- If you receive an error while installing (because your user is not administrator), change the "Install CARMEN for yourself, or anyone who uses this computer" setting to "Just me"
- Press "Next"
- Press "Next" again
- Once installed, you can run "CARMEN: Casting And Role Management Equality Network" from the start menu, under the "CARMEN" folder
If you have been given a username and password for a database server, you should connect as follows;
- Run "CARMEN: Casting And Role Management Equality Network" from the start menu
- Click the "Connect" button on the right hand side of the "CARMEN: Show Selection" window
- In the "Connect to Database" dialog, type the address of the server (eg. example.domain.com) in "Server Host"
- Type the name of the database (eg. show_2021) in "Database Name"
- Type your username and password in "Username" and "Password"
- Click "Connect" to save this connection, and open the show
- Next time you open CARMEN, you can skip steps 2-6 by selecting the saved connection from the list below the "New"/"Open"/"Connect" buttons
- Click "Auditioning" to go to the audition page.
- It is HIGHLY recommended that you tick "Auto-save on applicant change" so that changes are saved every time you switch between applicants, rather than only when you press the "Save" button to return to the main menu. Alternatively, you can press Ctrl+S at any time to save without returning to the main menu.
Once you have connected to a database server, or opened a local database file using the "New" or "Open" buttons, the main menu is shown. In the lower left quadrant of this window you will see the list of "steps" in the casting process.
Each step is either:
- ticked in green if complete,
- crossed in red if there are errors,
- or highlighted yellow if they are not yet done.
Under normal circumstances you would work on each of these steps in order, completing each before moving to the next. Clicking a step shows a sub page with the required information and controls to complete the step. A brief description of what you can do on each page is included below.
This page allows configuring the various parameters about how YOUR show structure works.
The audition criteria define what you will be marking each applicant on in the audition.
The cast group(s) is what successful applicants will be selected into. For a simple show, this may be a single group called "Cast". Successful applicants can only be in one cast group.
Alternative casts may be used if you use multiple sets of cast members, which alternate performances. Every role must be allocated to one person in each alternative cast.
Cast tags can be applied to any successful applicant (that is, anyone in any cast group) which meet certain criteria. Applicants may have more than one tag.
Section types define the types of structures you will have in your show, for example, Acts or Scenes. Each section type may have different rules about duplicate roles or being cast in consecutive items within that section.
Requirements define what attributes are needed by applicants to be eligible or desirable for a certain cast group, tag or role.
This page allows adding and editing applicant information. It will be marked as complete when enough appicants have been registered to meet the required number of cast members in each cast group, and every appicant has all basic information (name, age, gender) completed.
Applicants may have some audition marks pre-populated at this point, but it is not required for this step to be complete.
This page is similar to the Registering Applicants page, except that the goal has shifted to entering the applicant's audition marks for each criteria. This will only be marked as complete when every applicant has a mark for all required audition criteria.
This page allows selection of applicants into cast groups, alternative casts and tags. It can be done automatically based on the configured requirements, or manually, or a combination of the two.
It also creates the cast list, which involves assigning each cast member a unique number. If the cast member is in a cast group which uses alternative casts, there will be multiple cast members assigned the same number, one for each alternative cast. This represents that they are the matching cast member in the other alternative cast.
This page is where the structure of your show is configured. Starting from the heading of the "show", you can add as many items or sections recursively as you would like. The types of sections available are those configured in "Configuring Show".
This is also where you add the roles in the show, which will be allocated to cast members. Roles can be in 1 or more items, and items can be in any type of section, or directly in the "show". Roles cannot be directly in a section or the "show", therefore the simpliest structure would be a "show" with 1 item, containing 1 or more roles.
This page is where most of the time is spent in the casting process. Depending on the complexity of your show it can take many hours to allocate cast members to all of the roles which meet your various requirements and business logic.
On the left-hand side of this page you will see a list of all the roles in the show, separated by the item/section show structure as defined in "Configuring Items". Each item has either
- a green tick, to indicate that it is fully cast
- a red cross, to indicate that there is an error, or missing cast
- no icon, if the role has no cast allocated yet.
Each section/item can be expanded or collapsed to manage a long list of roles effectively. Below this list is the "Show completed" option which can be unchecked to hide any fully cast roles from the list, leaving only those which still have work to be done visible. There is also an overall progress bar.
Single clicking a role in this list will show the currently allocated cast and requirements of the role on the right-hand side of the screen. Double clicking the role name, the current cast list, or clicking the "Edit casting" button will show a list of available cast members for this role. In this list you can select or deselect cast members to allocate or de-allocate them from this role, and click "Save". There are also buttons to clear the cast for this role, or automatically allocate the most suitable cast.
Additionally, selecting an item or section in the roles list will show which roles within it have not yet been fully cast. There is also a button to allocate balanced cast to all (or a sub-selection) of these outstanding roles, which is both faster than automatically casting individual roles and also better balances the talent between the roles.
This step will be marked as complete when all roles have been cast, and all casting rules and requirements are met.
Behind the scenes, CARMEN uses specialised casting engines designed to give the best and most balanced outcomes. There are a number of variations of these engines which each operate in a slightly different way and may work better or worse in certain circumstances. These can be chosen in the settings, under Advanced.
The "Audition Engine" controls how the Applicant's overall ability number is calculated, as well as the raw suitability percentage for a given requirement. The NeuralAuditionEngine is used by default.
TODO
TODO
The "Selection Engine" controls how cast is accepted/rejected from the Cast Groups and Tags, as well as how the talent is balanced between Alternative Casts. The HybridPairsSatEngine is used by default.
TopPairsSatEngine is recommended for speed, BestPairsSatEngine is recommended for balancing the top talent, and ChunkedPairsSatEngine is recommended for balancing the whole list.
- For each cast group, list cast members in order by a criteria mark, and pair them down the list (skipping people who are in a same cast set with each other)
- Make sure each pair is in different casts
- Do this for each criteria and cast group
- If that fails (which it will, because the ideal solution never works), then increase the "chunk" size from 2 to 4 and try again (ie. make the lists and select chunks of 4, where 2 of each 4 must be in separate casts)
- Continue increasing chunk size by 2 until success
- This is a standard k-SAT problem and is solved using the DPLL algorithm
- Similar to chunked pairs, but rather than increase the chunk size, always keep it at 2 (ie. pairs) but reduce the total number of pairs by 1 each time until success
- This approach will more or less guarantee that the top cast members of each ability alternate, and cares less about those lower down each list
- Like top pairs, but once you have success with a certain number of pairs, rather than leave the rest of the lists as random, it chunks the rest of the list in a larger chunk size (ie. 4)
- Then reduce number of chunks of the new chunk size until success, then once again increase chunk size and try again
- This terminates when the chunk size is greater than the number of cast members at the bottom of list that aren't yet chunked
- Rather than pairs, put applicants down the list in sets of 3 and enforce that they can't ALL be in the same cast
- This is a much gentler requirement, which can stretch to the bottom of the list without needing a fallback
- However it doesn't strictly mandate that alternative casts must be even, therefore it its not good enough on its own.
- This makes it technically an SMT problem rather than pure SAT, as it has an external domain specific rule which must also be adhered to (that the casts are even)
- Therefore it is solved with the DPLL(T) algorithm (T stands for Theory or Test)
- This is more different again, because it uses the SAT rules for same cast sets, but then finds the optimum solution of the remaining options by trying to minimise a cost function
- In this instance the cost function is the difference between casts of a sum of ranks.
- Similarly to other approaches, you list the cast members in order for a certain criteria within a certain cast group. Ranks are given to these where 1 is the lowest and N is the highest
- Note that because we want equal marks to be equal ranks, N is not the number of applicants in a cast group but the number of unique marks for that criteria in that cast group
- This is done for each criteria, then the ranks are summed for each cast group, and the absolute difference between groups is calculated
- This is improved iteratively until all possible combinations have been tested (or proved to be worse than the current option) OR with the failsafe that its been more than 10 seconds without an improvement (because otherwise if you had 72 junior cast members and no same cast sets, there's a LOT of combinations to assess)
- This is another variation on SAT called weighted MAX-SAT and is solved with a Branch and Bound algorithm (because enumerating all options, and assessing each individually for the cost function would take literal years)
- This is very close to Top Pairs, but using the Branch & Bound solver from Rank Difference with the same cost function
- Ie. Pair applicants down the list, reduce number of pairs until any solution exists, then from the available solutions for that number of pairs, find the one with the lowest difference in rank sums between casts
- Allocate the same-cast-set applicants first, then filling the rest by sorting the applicants into cast number order
- This makes no attempt to balance the talent between the casts
The "Allocation Engine" controls how cast are allocated to Roles, and the order in which it recommends for you to cast the Roles. The RoleLearningAllocationEngine is used by default.
TODO
TODO
TODO
TODO
TODO