Periodic Reporting for period 1 - CAVAA (Counterfactual Assessment and Valuation for Awareness Architecture)
Reporting period: 2022-10-01 to 2023-09-30
The Counterfactual Assessment and Valuation for Awareness Architecture (CAVAA) project centers on the concept that awareness is vital for survival in a world of unseen elements and internal dynamics of agents and moral norms. This awareness combines perceptual data, memory, and inferred elements to form an internal virtual world. CAVAA aims to develop an integrated computational architecture to understand and engineer awareness in both biological and technological systems. The project will focus on perception, memory, virtualization, simulation, and integration, and apply these in robots and artificial agents across various use-cases like robot foraging, social robotics, and computer game benchmarks. These scenarios will test trade-offs such as efficiency versus robustness and gauge user acceptance.
The first year of the CAVAA project was characterized by both technical development and integration and deep theoretical discussions and conceptual refinement. A particularly significant achievement was the successful convergence of CAVAA partners during the DCBT 2023 Summer School at SRU, resulting in an intensive technical integration of the core computational models (WP1, WP2), setting the stage for subsequent developments and benchmarks (WP4). These collaborative sessions yielded a series of technical advances, with a special emphasis on the development, integration, and testing of the first modules of the CAVAA cognitive architecture. Core developments include the implementation of a reactive layer that emphasizes self-regulation, the enhancement of spatially-tuned features for efficient trajectory planning, and significant strides in episodic reinforcement learning, showcasing superior learning speed and memory storage. Furthermore, the first technical integration of various computational models, such as the Sequential Episodic Control with Model-Based RL, represents the first major step towards building CAVAA’s cognitive architecture. Work on WP3 also advanced the exploration of various robotic simulations, interface designs, and integration strategies. Key achievements include the clear definition of the types of information to be exchanged and the level of abstraction necessary for effective operation of these agents, facilitating the exchange of perception and motor control information between the robotic simulator and CAVAA’s computational models. Concurrently, a series of online scientific seminars and meetings held during the year culminated in the development of the first draft of a joint theoretical research paper, steered by UU, which worked to hone the project's theoretical underpinnings, offering clarity on consciousness and awareness definitions and their empirical validations. This manuscript serves as a foundational stone, linking theoretical postulations to real-world benchmarks, and thus paving the way for the project's subsequent phases. Collectively, these advancements not only reinforce the project's objectives but also exemplify the potent blend of theory, practice, and innovation that CAVAA champions.
Key developments in the first year demonstrate progress in developing the CAVAA cognitive architecture, with a focus on redefining decision-making and action selection in AI systems leveraging awareness. Notable achievements include the Whitened Sparse Autoencoder, converting perceptual states into discrete memory events, and the Sequential Episodic Control (SEC) model, advancing episodic reinforcement learning. SEC's sequential chaining of episodic memories enhances learning speed and memory capacity, setting a new benchmark in efficiency. Model-Based RL advancements, incorporating techniques like Prioritized Sweeping and Bidirectional Search, demonstrate adaptability and curiosity-driven learning in dynamic environments.
The integration of SEC and Model-Based RL further advances CAVAA's cognitive architecture, exemplifying flexible decision-making in dynamic scenarios. Theoretical progress, notably in the "Artificial Awareness" manuscript, differentiates 'consciousness' and 'awareness' and introduces consciousness profiles. This novel approach contributes to discussions on artificial awareness, offering new perspectives and validation methods.
The integration of SEC and Model-Based RL further advances CAVAA's cognitive architecture, exemplifying flexible decision-making in dynamic scenarios. Theoretical progress, notably in the "Artificial Awareness" manuscript, differentiates 'consciousness' and 'awareness' and introduces consciousness profiles. This novel approach contributes to discussions on artificial awareness, offering new perspectives and validation methods.