Agents' roles in our lives increasingly matter as they engage with people in a variety of important tasks. To achieve successful human-agent teamwork, it is critical to know the differences and similarities in people's attitudes towards human and agent teammates in virtual environments. It is unclear to what extent we can rely on the rich literature on interpersonal trust, i.e., trust between humans, while designing trustworthy agent teammates for human-agent teamwork and constructing hypotheses for human-agent trust research. This study empirically investigates the differences in the growth of human trust in and reliance on human and agent teammates during initial interactions. We developed a team coordination game, the Game of Trust, in which two players repeatedly cooperate to complete team tasks without prior assignment of subtasks. The effects of teammate type, i.e., human vs. agent, are evaluated by performing an extensive set of controlled experiments with participants recruited from Amazon Mechanical Turk. We collect both teamwork performance data as well as surveys to gauge participants' trust in their teammates. The empirical results show that humans' trust attitudes towards human and agent teammates differ: trust in and reliance on teammate and team performance were slightly higher when playing with the agent teammate. Moreover, the level of trustworthiness of a teammate is more influential on human trust compared to teammate type. These findings enhance our understanding of changes in human trust concerning teammate type towards achieving successful virtual teamwork.

We discuss an application of the INFINITE negotiation architecture for developing agents that can negotiate with others while representing its user's preferences. We developed an agent, Draft Agent, that was entered into the 2019 Human-Agent league (HAL) of the Autonomous Negotiating Agent Competition (ANAC). We discuss Draft Agent's performance, highlighting where it worked well and aspects that can be further improved. A key feature of Draft Agent is the use of an alternate-issue-selection protocol to model the opponent's preference structure. The learnt preferences are then used to propose a fair, and where possible, win-win deal. Though this approach allows Draft Agent to obtain relatively high individual as well as joint utility, it might be considered somewhat rigid by human users and hence scores comparatively low on the likeability scale. We present a detailed analysis of the comparative performance of Draft Agent and the competing finalists of the HAL competition. We also suggest some options to further improve Draft Agent's performance and likeability.

People use the knowledge acquired from past experiences in assessing the trustworthiness of a trustee. In a time where the agents are being increasingly accepted as partners in collaborative efforts and activities, it is critical to understand all aspects of human trust development in agent partners. For human-agent virtual ad hoc teams to be effective, humans must be able to trust their agent counterparts. To earn the humans’ trust, agents need to quickly develop an understanding of the expectation of human team members and adapt accordingly. This study empirically investigates the impact of past experience on human trust in and reliance on agent teammates. To do so, we developed a team coordination game, the Game of Trust (GoT), in which two players repeatedly cooperate to complete team tasks without prior assignment of subtasks. The effects of past experience on human trust are evaluated by performing an extensive set of controlled experiments with participants recruited from Amazon Mechanical Turk, a crowdsourcing marketplace. We collect both teamwork performance data as well as surveys to gauge participants’ trust in their agent teammates. The results show that positive (negative) past experience increases (decreases) human trust in agent teammates; lack of past experience leads to higher trust levels compared to positive past experience; positive (negative) past experience facilitates (hinders) reliance on agent teammates; the relationship between trust in and reliance on agent teammates is not always correlated. These findings provide clear and significant evidence of the influence of key factors on human trust in virtual agent teammates and enhance our understanding of the changes in human trust in peer-level agent teammates with respect to past experience.

Interacting with strangers and agents through computer networks has become a routine aspect of our daily lives. In such environments, reputation plays a critical role in determining our future interactions and satisfaction derived from them. This paper empirically investigates the effects of agent reputation on humantrust in andbehavior towards "peer'' level agent teammates over repeated interactions. We developed a team coordination game, the Game of Trust, in which a human player and an agent player repeatedly cooperate to complete team tasks without prior assignment of subtasks. Before the game begins, the agent player is introduced with either positive or negative reputation to the human player. The effects of agent reputation are evaluated by performing an extensive set of controlled experiments with participants recruited from Amazon Mechanical Turk, a crowdsourcing marketplace. We collect both teamwork performance data as well as surveys to gauge participants' trust in their agent teammates. The empirical results show that positive (negative) agent reputation led to greater (lower) human trust in agent teammates. Moreover, the interplay between the game expertise and expectation from agent teammate significantly affected the influence of reputation. These findings enhance our understanding of changes in human trust with respect to agent reputation towards achieving successful human-agent teamwork.

Social norms serve as an important mechanism to regulate the behaviours of agents and to facilitate coordination among them in multiagent systems. One important research question is how a norm can rapidly emerge through repeated local interaction within agent societies under different environments when their coordination space becomes large. To address this problem, we propose a hierarchically heuristic learning strategy (HHLS) under the hierarchical social learning framework. Subordinate agents report their information to their supervisors, while supervisors can generate instructions (rules and suggestions) based on the information collected from their subordinates. Subordinate agents heuristically update their strategies based on both their own experience and the instructions from their supervisors. Extensive experiment evaluations show that HHLS can support the emergence of desirable social norms more efficiently and can be applicable in a much wider range of multiagent interaction scenarios compared with previous work. The influence of key related factors (e.g., different topologies, population, neighbourhood and action space size, cluster size) are also investigated and new insights are obtained as well.

Tags and other characteristics, externally perceptible features that are consistent among groups of animals or humans, can be used by others to determine appropriate response strategies in societies. This usage of tags can be extended to artificial environments, where agents can significantly reduce cognitive effort spent on appropriate strategy choice and behaviour selection by reusing strategies for interacting with new partners based on their tags. Strategy selection mechanisms developed based on this idea have successfully evolved stable cooperation in games such as the Prisoner's Dilemma game but relies upon payoff sharing and matching methods that limit the applicability of the tag framework. Our goal is to develop a general classification and behaviour selection approach based on the tag framework. We propose and evaluate alternative tag matching and adaptation schemes for a new, incoming individual to select appropriate behaviour against any population member of an existing, stable society. Our proposed approach allows agents to evolve both the optimal tag for the environment as well as appropriate strategies for existing agent groups. We show that these mechanisms will allow for robust selection of optimal strategies by agents entering a stable society and analyse the various environments where this approach is effective.

Steven Brams's [(1994). Theory of moves. Cambridge University Press] Theory of Moves (TOM) is an alternative to traditional game theoretic treatment of real-life interactions, in which players choose strategies based on analysis of future moves and counter-moves that arise if game play commences at a specified start state and either player can choose to move first. In repeated play, players using TOM rationality arrive at nonmyopic equilibria. One advantage of TOM is its ability to model scenarios in which power asymmetries exist between players. In particular, threat power, i.e. the ability of one player to threaten and sustain immediate, globally disadvantageous outcomes to force a desirable result long term, can be utilised to induce Pareto optimal states in games such as Prisoner's Dilemma which result in Pareto-dominated outcomes using traditional methods. Unfortunately, prior work on TOM is limited by an assumption of complete information. This paper presents a mechanism that can be used by a player to utilise threat power when playing a strict, ordinal 2×2 game under incomplete information. We also analyse the benefits of threat power and support in this analysis with empirical evidence.

Theory of Moves (TOM) is a novel approach to game theory for determining rational strategies during the play of dynamic games [Brams, S J. (1994). Theory of moves. Cambridge, UK: Cambridge University Press]. While alternate models such as normal form games exist, players of these games are limited to single shot interactions with each other, but within TOM, sequences of moves and counter moves are allowed. As a consequence of this framework potential cyclic behaviour may arise. Unfortunately, standard TOM framework suggests that players do not move from the initial state if the possibility of cyclic behaviour is detected. However, in a plethora of real life scenarios, cycling can benefit a player over time. We first extend the TOM framework by allowing players to choose how much time to stay in each state while specifying time limits for moves. This generalisation allows for cycling behaviour in addition to normal, acyclic TOM play. We present additional rationality rules to handle the choice of move time and cyclic play and identify conditions for the existence of solutions that involve cycles. Moreover, if solutions do exist, equilibrium are determined so a player can predict the rational outcome upon engaging a cycle. A variety of time constraints on move times are investigated and the effects of these contrasts on the solution space and equilibrium are analysed.

Convention emergence solves the problem of choosing, in a decentralized way and among all equally beneficial conventions, the same convention for the entire population in the system for their own benefit. Our previous work has shown that reaching 100% agreement is not as straighforward as assumed by previous researchers, that, in order to save computational resources fixed the convergence rate to 90% (measuring the time it takes for 90% of the population to coordinate on the same action). In this article we present the notion of social instruments as a set of mechanisms that facilitate and accelerate the emergence of norms from repeated interactions between members of a society, only accessing local and public information and thus ensuring agents' privacy and anonymity. Specifically, we focus on two social instruments: rewiring and observation. Our main goal is to provide agents with tools that allow them to leverage their social network of interactions while effectively addressing coordination and learning problems, paying special attention to dissolving metastable subconventions. The first experimental results show that even with the usage of the proposed instruments, convergence is not accelerated or even obtained in irregular networks. This result leads us to perform an exhaustive analysis of irregular networks discovering what we have defined as Self-Reinforcing Structures (SRS). The SRS are topological configurations of nodes that promote the establishment and persistence of subconventions by producing a continuous reinforcing effect on the frontier agents. Finally, we propose a more sophisticated composed social instrument (observation + rewiring) for robust resolution of subconventions , which works by the dissolution of the stable frontiers caused by the Self-Reinforcing Substructures (SRS) within the social network.

We present the notion of Social Instruments as mechanisms that facilitate the emergence of conventions from repeated interactions between members of a society. Specifically, we focus on two social instruments: rewiring and observation. Our main goal is to provide agents with tools that allow them to leverage their social network of interactions when effectively addressing coordination and learning problems, paying special attention to dissolving metastable subconventions. Our initial experiments throw some light on how Self-Reinforcing Substructures (SRS) in the network prevent full convergence to society-wide conventions, resulting in reduced convergence rates. The use of an effective composed social instrument, observation + rewiring, allow agents to achieve convergence by eliminating the subconventions that otherwise remained meta-stable.

Remotely deployed sensor networks are vulnerable to both physical and electronic security breaches. The sensor nodes, once compromised, can send erroneous data to the base station, thereby possibly compromising network effectiveness. We assume that sensor nodes are organized in a hierarchy and use an offline neural network-based learning technique to predict the data sensed at any node given the data reported by its siblings in the hierarchy. This allows us to detect malicious nodes even when the siblings are not sensing data from the same distribution. The speed of detection of compromised nodes, however, critically depends on the mechanism used to update the reputation of the sensor nodes over time. We compare and contrast the relative strengths of a statistically grounded scheme and a reinforcement learning-based scheme both for their robustness to noise and responsiveness to change in sensor behavior. We first extend an existing mechanism to improve detection capability for smaller errors. Next we analyze the influence of different discount factors, including unweighted, exponential and linear discounts, on the tradeoff between responsiveness and robustness. We both develop a theoretical analysis to understand the tradeoff and perform experimental verification of our predictions by varying the patterns in sensed data.

When an agent learns in a multi-agent environment, the payoff it receives is dependent on the behaviour of the other agents. If the other agents are also learning, its reward distribution becomes non-stationary. This makes learning in multi-agent systems more difficult than single-agent learning. Prior attempts at value-function based learning in such domains have used off-policy Q-learning that do not scale well as the cornerstone, with restricted success. This paper studies on-policy modifications of such algorithms, with the promise of scalability and efficiency. In particular, it is proven that these hybrid techniques are guaranteed to converge to their desired fixed points under some restrictions. It is also shown, experimentally, that the new techniques can learn (from self-play) better policies than the previous algorithms (also in self-play) during some phases of the exploration.

We have evaluated the effectiveness of a probabilistic reciprocity scheme for promoting co-operation among self-interested agents. The probabilistic reciprocity mechanism is used to determine whether an agent should co-operate when approached for help by another agent. The situation becomes more complex when a group of agents seeks help from another group. The opinions of the members of the helping group about each of the asking group members can be combined to evaluate such a request for help. Exploitative agents would want to be part of groups that receive help from other groups, but will try to prevent its group from helping other groups. Such agents, revealing false opinion about the reputation of others, can cause unwarranted rejection of help requests from other groups. This leads to global performance degradation in terms of reduced inter-group co-operation and increased cost for the individual agents. We study the viability of reciprocative agents in randomly formed groups and when groups are formed by agents contracting other helpful agents. Group helping decisions are based on both average and worst combined ratings of group members. A key result from our study is that lying exploitative agents, who provide false opinions about other agents, become ineffective when focused group selection is enabled.

. Social agents, both human and computational, inhabiting a world containing multiple active agents, need to coordinate their activities. This is because agents share resources, and without proper coordination or ‘rules of the road’, everybody will be interfering with the plans of others. As such, we need coordination schemes that allow agents to effectively achieve local goals without adversely affecting the problem-solving capabilities of other agents. Researchers in the field of Distributed Artificial Intelligence (DAI) have developed a variety of coordination schemes under different assumptions about agent capabilities and relationships. Whereas some of these researchers have been motivated by human cognitive biases, others have approached it as an engineering problem of designing the most effective coordination architecture or protocol. We evaluate individual and concurrent learning by multiple, autonomous agents as a means for acquiring coordination knowledge. We show that a uniform reinforcement learning algorithm suffices as a coordination mechanism in both cooperative and adversarial situations. Using a number of multi-agent learning scenarios with both tight and loose coupling between agents and with immediate as well as delayed feedback, we demonstrate that agents can consistently develop effective policies to coordinate their actions without explicit information sharing. We demonstrate the viabilityof using both the Q-learning algorithm and genetic algorithm based classifier systems with different pay-off schemes, namely the bucket brigade algorithm (BBA) and the profit sharing plan (PSP), for developing agent coordination on two different multi-agent domains. In addition, we show that a semi-random scheme for action selection is preferable to the more traditional fitness proportionate selection scheme used in classifier systems.

Attribute based classification has been one of the most active areas of machine learning research over the past decade. We view the problem of hypotheses formation for classification as a search problem. Whereas previous research acquiring classification knowledge have used a deterministic bias for forming generalizations, we use a more random bias for taking inductive leaps. We re-formulate the supervised classification problem as a function optimization problem, the goal of which is to search for a hypotheses that minimizes the number of incorrect classifications of training instances. We use a simulated annealing based classifier (SAC) to optimize the hypotheses used for classification. The particular variation of simulated annealing algorithm that we have used is known as Very Fast Simulated Re-annealing (VFSR). We use a batch-incremental mode of learning to compare SAC with a genetic algorithm based classifier, GABIL, and a traditional incremental machine learning algorithm, ID5R. By using a set of artificial target concepts, we show that SAC performs better on more complex target concepts.>

Cooperative information agents need mechanisms that enable them to work together effectively while solving common problems. We investigate the use of commitment by agents to proposed actions as a mechanism that allow agents to work concurrently on interdependent problems. Judicious use of commitment can not only increase the throughput of cooperative information systems, but also allow them to deal flexibly with dynamically changing environments. We use the domain of distributed scheduling to demonstrate that static commitment strategies are ineffective. Results from simulated experiments are used to identify the environmental features on which an adaptive commitment strategy should be predicated.