2022 Fiscal Year Research-status Report
On progressing human understanding in the shadow of superhuman deep learning artificial intelligence entities
| Project/Area Number |
22K00015
|
|---|
| Research Institution | Akita International University |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2022-04-01 – 2025-03-31
|
| Keywords | artificial intelligence / machine learning / scientific method / philosophy of science / data mining / game of Go |
|---|
| Outline of Annual Research Achievements |
For our testbed application, the game of Go, we developed the `cost of passing' measure to assess the evolution of human understanding. It allows us to have a fine-grained context-dependent evaluation of playing performance and will be the primary tool in the historical game analysis. The cost of passing is also suitable for the visual 'fingerprinting' of a large number of games, which will help the data mining efforts greatly.
We also researched and evaluated several hardware options for computationally intensive bulk game analysis. We purchased, set up the equipment, and stress-tested it with machine learning algorithms for training better AIs for the analysis.
We also considered mathematically and philosophically what the existence of a perfect player would mean for this application domain. Philosophically it gives an interesting perspective on scientific realism, as the ultimate reality of a board game is the complete game tree, and it has a not-efficiently computable but seemingly straightforward existence. Mathematically, considering the space of possibilities for AI and human knowledge growth led to a side project for constructing transducers for representing perfect game strategies using logic programming.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
The most critical requirement for the project was the development of a suitable context-sensitive measure for playing strength, and this process went without any problems.
Designing an energy-efficient computing server setup for the computational analysis was challenging due to the uncertainties and delays in the hardware industry (repercussions of the former supply chain issues and the market changes). However, the delayed starting of the historical game analysis is likely to be offset by the advances in the analysis software (more efficient neural network architectures leading to faster game analysis).
Also adding to the progress is the unplanned work on constructing transducers from input-output pairs (lossless machine learning - metaphorically speaking). This long-term subproject will add to the philosophical discussion about the future of AIs and human understanding.
|
| Strategy for Future Research Activity |
1. The historical game analysis. The hardware purchase and the development of the new measure (the cost of passing) finished successfully in the first financial year. Before starting the bulk analysis, we will calibrate the software setup (number of visits, network choice, hardware optimization, fixing storage data formats, and process automation). The automated analysis will continue until the very end of the project.
2. Visualization of analysis results: extending the current software package (LambdaGo) to deal with the bulk analysis outputs (e.g., visualize the cost of passing fingerprints and other measures, creating summarizing statistics). This requires changing the visualization software packages.
3. Best practices survey. We also need to determine the current standard practices in utilizing AI knowledge. This will involve research visits (e.g., PI to Nihon Ki-in in Tokyo and RC to Akita International University).
4. Writing and presenting papers on the critical assessment of our proposed method of adopting a scientific approach in the learning process; on the ontology and mathematics of Go knowledge.
|
| Causes of Carryover |
Uncertainties in the computing industry delayed our purchases and prevented us from using the full amount of funding.
Most of the budget will be used for research visits (e.g. PI to Nihon-Kiin, the research collaborator (Antti Tormanen) to AIU, including his remuneration) and conference costs (travel, registration). Conference trips are needed for the dissemination of our results and interacting with other researchers to get feedback on our project. The visits of the research collaborator (professional Go player) are needed for intensive work on the Go-related part of the project. The visit of Nihon-Kiin by the PI is for surveying the best practices of AI usage by professional players.
Small amount is planned for purchasing books that may be needed in developing the philosophical arguments for using the scientific method for working with AIs.
No computing hardware purchase is needed any more, unless in the case of an emergency of one of the computing units breaking down. For the historical analysis, data storage devices/media will be required, since the analysis generates a huge amount of data, and the archival of those computational results should be off the live computing servers.
|
| Remarks |
The above Git repositories are the software packages that contain the tools used for this research project. They are publicly available and the development process is thoroughly documented there.
|
