An orientation-dependent optic-fiber accelerometer based on the excessively tilted fiber grating (ExTFG) inscribed in SM28 fiber is demonstrated, which is based on the optical power demodulation scheme. Without any complicated processing, the cladding mode resonances of the bare ExTFG show high sensitivity to slight perturbation of bending. Due to its excellent azimuth-related bending properties, such a bare ExTFG fixed on a simple cantilever beam has exhibited strong orientation-dependent vibration properties. The experimental results show that a TE mode of the sensor can provide a maximum acceleration sensitivity of 74.14 mV/g at 72 Hz and maximum orientation sensitivity of 9.1 mV/deg while, for a TM mode, a maximum acceleration sensitivity of 57.85 mV/g at 72 Hz and maximum orientation sensitivity of 7.4 mV/deg could be achieved. These unique properties enable the sensor to act as a vector accelerometer for applications in many vibration measurement fields.
Non-transitive environments complicate the notion of a single ''best'' strategy: performance depends on the opponent population, and rankings are meaningful only relative to a specified opponent pool and protocol. We present a reproducible multi-agent benchmark for iterated Rock--Paper--Scissors and evaluate 54 agents from 18 archetypes---deep recurrent and transformer sequence models, actor--critic reinforcement learners, Bayesian/Markov predictors, classical classifiers, and rule-based baselines---in 500-round double round-robin tournaments across 10 random seeds. To support auditability, we define a simple regret certificate: a Lipschitz-type inequality that upper-bounds a best-response payoff gap using the $\ell_1$ discrepancy between an agent's predicted action distribution and an empirical estimate of the opponent's action distribution, computable online from logged predictions. Our experiments indicate that (i) recurrent predictors tend to achieve the highest and most stable scores, primarily by exploiting predictable opponents; (ii) rankings shift notably with the opponent pool (Spearman $\rho = 0.65$ between two evaluation rosters), with the top-ranked method changing across configurations; and (iii) the induced meta-game exhibits substantial non-transitivity, including 177 detected three-cycles in the pairwise payoff matrix. Under our 500-round online update budget and short-context design, transformer agents are competitive but do not outperform tuned recurrent baselines, which may reflect an inductive-bias mismatch in short-horizon adversarial play. Our code and analysis pipeline provide an extensible testbed for studying population-dependent evaluation and learning dynamics in canonical non-transitive games.
This paper explores human behavior in virtual networked communities, specifically individuals or groups' potential and expressive capacity to respond to internal and external stimuli, with assortative matching as a typical example. A modeling approach based on Multi-Agent Reinforcement Learning (MARL) is proposed, adding a multi-head attention function to the A3C algorithm to enhance learning effectiveness. This approach simulates human behavior in certain scenarios through various environmental parameter settings and agent action strategies. In our experiment, reinforcement learning is employed to serve specific agents that learn from environment status and competitor behaviors, optimizing strategies to achieve better results. The simulation includes individual and group levels, displaying possible paths to forming competitive advantages. This modeling approach provides a means for further analysis of the evolutionary dynamics of human behavior, communities, and organizations in various socioeconomic issues.