Continual learning (CL) is a new online learning technique over sequentially generated streaming data from different tasks, aiming to maintain a small forgetting loss on previously-learned tasks. Existing work focuses on reducing the forgetting loss under a given task sequence. However, if similar tasks continuously appear to the end time, the forgetting loss is still huge on prior distinct tasks. In practical IoT networks, an autonomous vehicle to sample data and learn different tasks can route and alter the order of task pattern at increased travelling cost. To our best knowledge, we are the first to study how to opportunistically route the testing object and alter the task sequence in CL. We formulate a new optimization problem and prove it NP-hard. We propose a polynomial-time algorithm to achieve approximation ratios of $\frac{3}{2}$ for underparameterized case and $\frac{3}{2} + {r^{1 - T}}$ for overparameterized case, respectively. Simulation results verify our algorithm’s close-to-optimum performance.

Reinforcement learning from human feedback (RLHF) has become an essential step in fine-tuning large language models (LLMs) to align them with human preferences. However, human labelers are selfish and have diverse preferences. They may strategically misreport their online feedback to influence the system’s aggregation towards their own preferences. Current practice simply averages labelers’ feedback per time and fails to identify the most accurate human labeler, leading to linear regret ${\mathcal{O}}(T)$ for T time slots. To our best knowledge, we are the first to study online learning mechanisms against strategic human labelers in the LLM fine-tuning process. We formulate a new dynamic Bayesian game and dynamically adjust human labelers’ weights in the preference aggregation, ensuring their truthful feedback and sublinear regret ${\mathcal{O}}\left({{T^{1/2}}}\right)$. Simulation results demonstrate our mechanism’s great advantages over the existing benchmark schemes.

The binary exponential backoff scheme is widely used in WiFi 7 and still incurs poor throughput performance under dynamic channel environments. Recent model-based approaches (e.g., non-persistent and p-persistent CSMA) simply optimize backoff strategies under a known and fixed node density, still leading to a large throughput loss due to inaccurate node density estimation. This paper is the first to propose LLM transformer-based in-context learning (ICL) theory for optimizing channel access. We design a transformer-based ICL optimizer to pre-collect collision-threshold data examples and a query collision case. They are constructed as a prompt as the input for the transformer to learn the pattern, which then generates a predicted contention window threshold (CWT). To train the transformer for effective ICL, we develop an efficient algorithm and guarantee a near-optimal CWT prediction within limited training steps. As it may be hard to gather perfect data examples for ICL in practice, we further extend to allow erroneous data input in the prompt. We prove that our optimizer maintains minimal prediction and throughput deviations from the optimal values. Experimental results on NS-3 further demonstrate our approach's advantage.

Today mobile crowdsourcing platforms invite users to provide anonymous reviews about service experiences, yet many reviews are found biased to be extremely positive or negative. The existing methods find it difficult to learn from biased reviews to infer the actual service state, as the state can also be extreme and the platform cannot verify the truthfulness of reviews immediately. Further, reviewers can hide their (positive or negative) bias types and proactively adjust their anonymous reviews against the platform's inference. To our best knowledge, we are the first to study how to save mobile crowdsourcing from cheap-talk and strategically learn from biased users' reviews. We formulate the problem as a dynamic Bayesian game, including users' service-type messaging and the platform's follow-up rating/inference. Our closed-form PBE shows that an extremely-biased user may still honestly message to convince the platform of listening to his review. Such Bayesian game-theoretic learning obviously outperforms the latest common schemes especially when there are multiple diversely-biased users to compete. For the challenging single-user case, we further propose a time-evolving mechanism with the platform's commitment inferences to ensure the biased user's truthful messaging all the time, whose performance improves with more time periods to learn from more historical data.

Today many users are invited by a crowdsourcing platform (e.g., TripAdvisor, and Waze) to provide their anonymous reviews about service experiences (e.g., in hotels, restaurants, and trips), yet many reviews are found biased to be extremely positive or negative. It is difficult to learn from biased users' reviews to infer actual service state, as the service state can also be extreme and the platform cannot verify immediately. Further, due to the anonymity, reviewers can hide their bias types (positive or negative) from the platform and adaptively adjust their reviews against the platform's inference. To our best knowledge, we are the first to study how to save crowdsourcing from cheap-talk and strategically learn the actual service state from biased users' reviews. We formulate the problem as a dynamic Bayesian game, including the unknown users' messaging and the platform's follow-up inference. Through involved analysis, we provide closed-form expressions for the Perfect Bayesian Equilibrium (PBE). Our PBE shows that the platform's strategic learning can successfully prevent biased users from cheap-talk in most cases, where a user even with extreme bias still honestly messages to convince the platform of listening to his review. We prove that the price of anarchy (PoA) is 2, telling that the social cost can at most be doubled in the worst-case. As the user number becomes large, our platform always learns the actual service state. Perhaps surprisingly, the platform's expected cost may be worse off after adding one more user.

After upgrading to 5G, a network operator still faces congestion when providing the ubiquitous wireless service to the crowd. To meet users' ever-increasing demand, some other operators (e.g., Fon) have been developing another crowdsourced WiFi network to combine many users' home WiFi access points and provide enlarged WiFi coverage to them. While the 5G network experiences negative network externality, the crowdsourced WiFi network helps offload traffic from 5G and its service coverage exhibits positive externality with its subscription number. To our best knowledge, we are the first to investigate how these two heterogeneous networks of diverse network externalities co-exist from an economic perspective. We propose a dynamic game theoretic model to analyze the hybrid interaction among the 5G operator, the crowdsourced WiFi operator, and users. Our user choice model with WiFi's complementarity for 5G allows users to choose both services, departing from the traditional economics literature where a user chooses one over another alternative. Despite of non-convexity of the operators pricing problems, we prove that the 5G operator facing severe congestion may lower his price, and he benefits from the introduction of crowdsourced WiFi. However, 5G operator with mild congestion charges users more and all the users' payoffs may decrease.

Online content platforms are concerned about the freshness of their content updates to their end customers, and increasingly more platforms now invite and pay the crowd to sample real-time information (e.g., traffic observations and sensor data) to help reduce their ages of information (AoI). How much crowdsourced data to sample and buy over time is a critical question for a platform's AoI management, requiring a good balance between its AoI and the incurred sampling cost. This question becomes more interesting by considering the stage after sampling, where multiple platforms coexist in sharing the content delivery network of limited bandwidth, and one platform's update may jam or preempt the others' under negative network externalities. When these selfish platforms know each other's sampling cost, we formulate their competition as a non-cooperative game and show they want to over-sample to reduce their own AoIs, causing the price of anarchy (PoA) to be infinity. To remedy this huge efficiency loss, we propose a trigger mechanism of non-monetary punishment in a repeated game to enforce the platforms' cooperation to approach the social optimum. We also study the more challenging scenario of incomplete information that some new platform hides its private sampling cost information from the other incumbent platforms in the Bayesian game. Perhaps surprisingly, we show that even the platform with more information may get hurt. We successfully redesign the trigger-and-punishment mechanism to negate the platform's information advantage and ensure no cheating. Our extensive simulations show that the mechanisms can remedy the huge efficiency loss due to platform competition, and the performance improves as we have more incumbent platforms with known cost information.

Online content platforms are concerned about the freshness of their content updates to their end customers, and increasingly more platforms now invite and pay the crowd to share real-time information (e.g., news and sensor data) to help reduce their ages of information (AoI). How much crowdsourced data to sample and buy over time is a critical question for a platform's AoI management, requiring a good balance between its AoI and the incurred sampling cost. This question becomes more interesting by considering the stage after sampling, where two platforms coexist in sharing the content delivery network of limited bandwidth, and one platform's update may jam or preempt the other's under negative network externalities. When the two selfish platforms know each other's sampling costs, we formulate their interaction as a non-cooperative game and show both want to over-sample to reduce their own AoI, causing the price of anarchy (PoA) to be infinity. To remedy this huge efficiency loss, we propose a non-monetary trigger mechanism of punishment in a repeated game to enforce the platforms' cooperation to achieve the social optimum. We also study the more challenging incomplete information scenario that platform 1 knows more information about sampling cost than platform 2 by hiding its sampling cost information in the Bayesian game. Perhaps surprisingly, we show that even platform 1 may get hurt by knowing more information. We successfully redesign the trigger-and-punishment mechanism to negate platform 1's information advantage and ensure no cheating. As compared to the social optimum, extensive simulations show that the mechanisms can remedy the huge efficiency loss due to platform competition in different information scenarios.