{
  "File Number": "1024",
  "Title": "Collaborating with Humans without Human Data",
  "Limitation": "Limitations and future work Our method currently relies on the manual process of initially training and selecting a diverse set of partners. This is not only time consuming, but also prone to researcher biases that may negatively influence the behavior of the created agents. Additionally, while we found FCP with a partner population size of N = 32 sufficient here, for more complex games, FCP may require an unrealistically large partner population size to represent sufficiently diverse strategies. To address these concerns, methods for automatically generating partner diversity for common-payoff games may be important. Possibilities include adaptive population matchmaking as been used in competitive zero-sum games [69], as well as auxiliary objectives that explicitly encourage behavioral diversity [19, 45, 46]. Our method requires a known and fixed reward function. We also focus on one domain in order to compare with prior work which has argued that human-in-the-loop training is necessary. Consequently, the resulting agents are only designed to adaptively collaborate on a single task, and not to infer human preferences in general [1, 33, 59]. Moreover, if a task’s reward function is poorly aligned with how humans approach the task, our method may well produce subpar partners, as would any method without access to human data. Thus, additional domains and tasks should be studied to better understand how our method generalizes. Targeted experiments to test specific forms of generalization may be especially helpful in this regard [38], as could approaches that procedurally generate environment layouts requiring diverse solutions [22]. Finally, it may be possible to produce even stronger agent assistants by combining the strengths of FCP (i.e. diversity) and BCP (i.e. human-like play). Indeed, Knott et al. [38] recently demonstrated that modifying BCP to train with multiple BC partners produces more robust collaboration with held-out agents, a finding that would be interesting to test with human partners. Societal impact A challenge for this line of work is ensuring agent behavior is aligned with human values (i.e. the AI value alignment problem [23, 59]). Our method has no guarantees that the resulting policy aligns with the preferences, intentions, or welfare of its potential partners. It likewise does not exclude the possibility that the target being optimized for is harmful (e.g. if the agent’s partner expresses preferences or intentions to harm others). This could therefore produce negative societal effects either if training leads to poor alignment or if agents are optimized for harmful metrics. One potential strategy for mitigating these risks is the use of human preference data [15]. Such data could be used to fine-tune and filter trained agents before deployment, encouraging better alignment with human values. A key question in this line of research is how human preference data should be\naggregated—or selected, in the case of expert preferences—when our aim is to create socially aligned agents (i.e. agents that are sufficiently aligned for everyone). Relatedly, targeted research on human beliefs and perceptions of AI [48], and how they steer human-agent interaction, would help inform agent design for positive societal impact. For instance, developers could incorporate specific priors into agents to reinforce tendencies for fair outcomes [20, 32]. Conclusion We proposed a method which is both effective at collaborating with humans and simple to implement. We also presented a rigorous and general methodology for evaluating with humans and eliciting their preferences. Together, these establish a strong foundation for future research on the important challenge of human-agent collaboration for benefiting society.",
  "Reviewer Comment": "Reviewer_1: Quality: This paper makes possible the zero-shot coordination with humans. Yet, this paper does not give a theoretical basis for this, nor does it give a proof. The experimental environment to prove the effectiveness of this method is relatively simple.\nClarity: This paper illustrates the idea and logic in a very clear and easy-to-understand way.\nOriginality: The main innovation of this paper is to use the checkpoint in the training process, so as to improve the diversity of the behavior of the collaborators who train the collaborator agent, and make the collaborator agent have strong generalization ability. The application of this method in the field of agent human cooperation is an innovation, but in the field of deep computer vision, there has been the realization of similar idea.\nSignificance: This paper is of great significance to the research field of human agent cooperation. It proposes that human data can be unnecessary for training agents which collaborate with humans. This attempt is of significance.\nThe main weakness of this paper is that the author does not give the theoretical connection between the method and the effect, and does not give a large number of results under different experimental environment.\nYet, it is an interesting problem to train agents that can cooperate well with human beings without using human data. The author's attempt is worthy of affirmation.\nLimitations And Societal Impact:\nQuestion: 1.Can the authors give theoretical proof of what the thought is effective? 2.Can the experiment be carried out in more experimental environments to obtain more convincing experimental results?\nEthical Concerns:\nN/A\nNeeds Ethics Review: No\nTime Spent Reviewing: 5\n\nReviewer_2: The idea is strikingly simple and presented in an easy to follow manner. Furthermore, it shows significant gains when compared to prior work that relies simply on self-play or examples of optimal human behavior. The ablations due a good job justifying the use of previous model checkpoints, as well as the network architectures changes made v.s. prior work. Finally, the human study convincingly shows that this agent excels (v.s. baselines) at its stated goal - collaborating with humans.\nLimitations And Societal Impact:\nThe authors do a good job discussing their method's limitations as well as the potential negative social implications. However, I believe they should also mention that the computational/time burden of training N separate agents could prove overwhelming in more complicated settings than grid world (e.g. real world robotics, Atari, etc.). Future work should definitely focus on more efficiently learning a diverse agent set.\nEthical Concerns:\nI believe this work is exploratory and thus has limited ethical concerns at this time. The authors did a good job discussing potential long term side effects.\nNeeds Ethics Review: No\nTime Spent Reviewing: 1\n\nReviewer_3: This paper is well-written and timely given growing interest in applying AI to work with humans. The results show that FCP results in more robust models and not having to use collected human data is a key advantage. I can see others adopting this technique for their own domains. Although I am generally supportive of the work, I have concerns about the novelty and generalizability of the results.\nFirst, using prior agent \"checkpoint\" models to capture different skill levels is similar to previous work [Tylkin, Radanovic & Parkes, 2020]. There, the domain is different (Atari games) and the goal was to train a helper agent, but there are similarities in motivation and technique. Perhaps the authors can clarify the key differences? Also, the idea that diversity is important for training agents to collaborate is not exactly new; there is also prior work in using hand-coded policies or past teammate models, e.g., PLASTIC [Barrett, Rosenfeld, Kraus, Stone. 2017].\nI like the fact that different layouts were engineered to test different collaborative behaviors and robustness to different problem settings. However, it remains unclear how well the findings will generalize to other domains (e.g., Starcraft, Hanabi). A more thorough analysis would help support the main claims and to provide more insights. I also have some specific questions about the experiments:\nit would help to quantify how different the held-out agent populations are. Although the model structures may be different, the policies could be functionally very similar for this task?\nFor the experiment with humans, there appear to be performance differences on specific layouts? Figure 20 in the Appendix shows that on the asymmetric and forced layouts, FCP's performance is fairly similar to BCP and PP, respectively. Some elaboration is needed here, especially given the claim that \"FCP coordinates best with humans, achieving the highest scores across all maps\". Perhaps I am missing something here?\nIn Fig 8, why are the BCP metrics for pot preference so dissimilar from the human data? Presumably, BCP was trained using human data and would be able to replicate the preference distribution?\nOverall, this work contributes an interesting study and the findings (if generally applicable) are meaningful for human-AI collaboration. The method is simple to implement and works well. That said, I am on the fence given my concerns above. If the authors can address them, I am inclined to increase my score.\nPost-Reponse Comments\nThank you for the clear response. I've read through the other reviews and responses. Overall, I am in favor of the topic and questions, but I find the proposed method/approach lacks clear novelty. Training the model with multiple seeds and for different skill levels appears to be an incremental change. Stronger experiments with analysis into the policy difference/disagreement scores would help. While these scores are \"low-level\", they can help to at least indicate that the policies are different to some degree. As such, I am keeping my score.\nLimitations And Societal Impact:\nLimitations and societal impact are adequately discussed.\nNeeds Ethics Review: No\nTime Spent Reviewing: 3",
  "Limitations_refined": "Limitations and future work Our method currently relies on the manual process of initially training and selecting a diverse set of partners. This is not only time consuming, but also prone to researcher biases that may negatively influence the behavior of the created agents. Additionally, while we found FCP with a partner population size of N = 32 sufficient here, for more complex games, FCP may require an unrealistically large partner population size to represent sufficiently diverse strategies. To address these concerns, methods for automatically generating partner diversity for common-payoff games may be important. Possibilities include adaptive population matchmaking as been used in competitive zero-sum games [69], as well as auxiliary objectives that explicitly encourage behavioral diversity [19, 45, 46]. Our method requires a known and fixed reward function. We also focus on one domain in order to compare with prior work which has argued that human-in-the-loop training is necessary. Consequently, the resulting agents are only designed to adaptively collaborate on a single task, and not to infer human preferences in general [1, 33, 59]. Moreover, if a task’s reward function is poorly aligned with how humans approach the task, our method may well produce subpar partners, as would any method without access to human data. Thus, additional domains and tasks should be studied to better understand how our method generalizes. Targeted experiments to test specific forms of generalization may be especially helpful in this regard [38], as could approaches that procedurally generate environment layouts requiring diverse solutions [22]. Finally, it may be possible to produce even stronger agent assistants by combining the strengths of FCP (i.e. diversity) and BCP (i.e. human-like play). Indeed, Knott et al. [38] recently demonstrated that modifying BCP to train with multiple BC partners produces more robust collaboration with held-out agents, a finding that would be interesting to test with human partners. Societal impact A challenge for this line of work is ensuring agent behavior is aligned with human values (i.e. the AI value alignment problem [23, 59]). Our method has no guarantees that the resulting policy aligns with the preferences, intentions, or welfare of its potential partners. It likewise does not exclude the possibility that the target being optimized for is harmful (e.g. if the agent’s partner expresses preferences or intentions to harm others). This could therefore produce negative societal effects either if training leads to poor alignment or if agents are optimized for harmful metrics. One potential strategy for mitigating these risks is the use of human preference data [15]. Such data could be used to fine-tune and filter trained agents before deployment, encouraging better alignment with human values. A key question in this line of research is how human preference data should be\naggregated—or selected, in the case of expert preferences—when our aim is to create socially aligned agents (i.e. agents that are sufficiently aligned for everyone). Relatedly, targeted research on human beliefs and perceptions of AI [48], and how they steer human-agent interaction, would help inform agent design for positive societal impact. For instance, developers could incorporate specific priors into agents to reinforce tendencies for fair outcomes [20, 32].",
  "abstractText": "Collaborating with humans requires rapidly adapting to their individual strengths, weaknesses, and preferences. Unfortunately, most standard multi-agent reinforcement learning techniques, such as self-play (SP) or population play (PP), produce agents that overfit to their training partners and do not generalize well to humans. Alternatively, researchers can collect human data, train a human model using behavioral cloning, and then use that model to train “human-aware” agents (“behavioral cloning play”, or BCP). While such an approach can improve the generalization of agents to new human co-players, it involves the onerous and expensive step of collecting large amounts of human data first. Here, we study the problem of how to train agents that collaborate well with human partners without using human data. We argue that the crux of the problem is to produce a diverse set of training partners. Drawing inspiration from successful multi-agent approaches in competitive domains, we find that a surprisingly simple approach is highly effective. We train our agent partner as the best response to a population of self-play agents and their past checkpoints taken throughout training, a method we call Fictitious Co-Play (FCP). Our experiments focus on a two-player collaborative cooking simulator that has recently been proposed as a challenge problem for coordination with humans. We find that FCP agents score significantly higher than SP, PP, and BCP when paired with novel agent and human partners. Furthermore, humans also report a strong subjective preference to partnering with FCP agents over all baselines.",
  "1 Introduction": "Generating agents which collaborate with novel partners is a longstanding challenge for Artificial Intelligence (AI) [4, 16, 37, 52]. Achieving ad-hoc, zero-shot coordination [31, 66] is especially important in situations where an AI must generalize to novel human partners [6, 61]. Many successful approaches have employed human models, either constructed explicitly [14, 35, 53] or learnt implicitly [12, 60]. By contrast, recent work in competitive domains has shown that it is possible to reach humanlevel using model-free reinforcement learning (RL) without human data, via self-play [8, 9, 63, 64]. This begs the question: Can model-free RL without human data generate agents that can collaborate with novel humans?\nWe seek an answer to this question in the space of common-payoff games, where all agents work towards a shared goal and receive the same reward. Self-play (SP), in which an agent learns from repeated games played against copies of itself, does not produce agents that generalize well to novel co-players [10, 11, 21, 44]. Intuitively, this is because agents trained in self-play only ever need to coordinate with themselves, and so make for brittle and stubborn collaborators with new partners who act differently. Population play (PP) trains a population of agents, all of whom interact with each other [39]. While PP can generate agents capable of cooperation with humans in competitive team games [34], it still fails to produce robust partners for novel humans in pure common-payoff settings\n⇤Equal contribution.\n35th Conference on Neural Information Processing Systems (NeurIPS 2021).\n[12]. PP in common-payoff settings naturally encourages agents to play the same way, reducing strategic diversity and producing agents not so different from self-play [24].\nOur approach starts with the intuition that the key to producing robust agent collaborators is exposure to diverse training partners. We find that a surprisingly simple strategy is effective in generating sufficient diversity. We train N self-play agents varying only their random seed for neural network initialization. Periodically during training, we save agent “checkpoints” representing their strategy at that point in time. Then, we train an agent partner as the best-response to both the fully-trained agents and their past checkpoints. The different checkpoints simulate different skill levels, and the different random seeds simulate breaking symmetries in different ways. We refer to this agent training procedure as Fictitious Co-Play (FCP) for its relationship to fictitious self-play [7, 27, 28, 69]. We evaluate FCP in a fully-observable two-player common-payoff collaborative cooking simulator. Based on the game Overcooked [25], it has recently been proposed as a coordination challenge for AI [12, 50, 70]. State-of-the-art performance in producing agents capable of generalization to novel humans was achieved in [12] via behavioral cloning (BC) of human data. More precisely, BC was used to produce models that can stand in as human proxies during training in simulation, a method we call behavioral cloning play (BCP). We demonstrate that FCP outperforms BCP in generalizing to both novel agent and human partners, and that humans express a significant preference for partnering with FCP over BCP. Our method avoids the cost and potential privacy concerns of collecting human data for training, while achieving better outcomes for humans at test time.\nWe summarize the novel contributions of this paper as follows:\n1. We propose Fictitious Co-Play (FCP) to train agents capable of zero-shot coordination with humans (Section 2.1).\n2. We demonstrate that FCP agents generalize better than SP, PP, and BCP in zero-shot coordination with a variety of held-out agents (Section 4.2).\n3. We propose a rigorous human-agent interaction study with behavioral analysis and participant feedback (Section 5.1).\n4. We demonstrate that FCP significantly outperforms the BCP state-of-the-art, both in task score and in human partner preference (Section 5.2).",
  "2.1 Fictitious Co-Play (FCP)": "Diverse training conditions have been shown to make agents more robust, from environmental variations (i.e. domain randomization [54, 56, 67]) to heterogeneity in training partners [69]. We seek to train agents that are robust partners for humans in common-payoff games, and so extend this line of work to that setting.\nOne important challenge in collaborating with novel partners is dealing with symmetries [31]. For example, two agents A and B facing each other may move past each other by A going left and B going right, or vice versa. Both are valid solutions, but a good agent partner will adaptively switch between\nthese conventions if a human clearly prefers one over the other. A second important challenge is dealing with variations in skill level. Good agent partners should be able to assist both highly-skilled partners, as well as partners who are still learning.\nFictitious co-play (FCP) is a simple two-stage approach for training agents that overcomes both of these challenges (Figure 2, right). In the first stage, we train a diverse pool of partners. To allow the pool to represent different symmetry breaking conventions, we train N partner agents in self-play. Since these partners are trained independently, they can arrive at different arbitrary conventions for breaking symmetries. To allow the pool to represent different skill levels, we use multiple checkpoints of each self-play partner throughout training. The final checkpoint represents a fully-trained “skillful” partner, while earlier checkpoints represent less skilled partners. Notably, by using multiple checkpoints per partner, this additional diversity in skill incurs no extra training cost.\nIn the second stage, we train an FCP agent as the best response to the pool of diverse partners created in the first stage. Importantly, the partner parameters are frozen and thus FCP must learn to adapt to partners, rather than expect partners to adapt to it. In this way, FCP agents are prepared to follow the lead of human partners, and learn a general policy across a range of strategies and skills. We call our method “fictitious” co-play for its relationship to fictitious self-play in which competitive agents are trained with past checkpoints (in that case, to avoid strategy cycling) [7, 27, 28, 39, 69].",
  "2.2 Baselines and ablations": "We compare FCP agents to the three baseline training methods listed below, each varying only in their set of training partners, with the RL algorithm and architecture consistent across all agents:\n1. Self-play (SP), where agents learn solely through interaction with themselves. 2. Population-play (PP), where a population of agents are co-trained through random pairings. 3. Behavioral cloning play (BCP), where an agent is trained with a BC model of a human [12].\nWe also evaluate three variations on FCP to better understand the conditions for its success:\n1. To test the importance of including past checkpoints in training, we evaluate an ablation of FCP in which agents are trained only with the converged checkpoints of their partners (FCP T for “FCP minus time”).\n2. To test whether FCP would benefit from additional diversity in its partner population, we evaluate an augmentation of FCP in which the population of SP partners varies not just in random seed, but also in architecture (FCP+A for “FCP plus architectural variation”).\n3. To test whether architectural variation can serve as a full replacement for playing with past checkpoints, we evaluate the combination of both modifications (FCP T,+A).",
  "2.3 Environment": "Following prior work on zero-shot coordination in human-agent interaction, we study the Overcooked environment (see Figure 3) [12, 13, 38, 50, 70]. We draw particular inspiration from the environment in Carroll et al. [12]. For full details, see Appendix A.\nIn this environment, players are placed into a gridworld kitchen as chefs and tasked with delivering as many cooked dishes of tomato soup as possible within an episode. This involves a series of sequential high-level actions to which both players can contribute: collecting tomatoes, depositing them into cooking pots, letting the tomatoes cook into soup, collecting a dish, getting the soup, and delivering it. Upon a successful delivery, both players are rewarded equally.\nTo effectively complete the task, players must learn to navigate the kitchen and interact with objects in the correct order, all while maintaining awareness of their partner’s behavior to coordinate with them. This environment therefore presents the challenges of both movement and strategic coordination.\nEach player observes an egocentric RGB view of the world, and at every step can perform one of six actions: stand still, move {up, down, left, right}, interact. The behavior of interact varies based on the cell which the player is facing (e.g. place tomato on counter).",
  "2.4 Implementation details": "Here we highlight several key implementation details for our training methods. For full details, including the architectures, hyperparameters, and compute used, please see Appendix B.\nFor our reinforcement learning agents, we use the V-MPO [65] algorithm along with a ResNet [26] plus LSTM [29] architecture which we found led to optimal behavior across all layouts. Agents are trained using a distributed set of environments running in parallel [17], each sampling two agents from the training population to play together every episode.\nBoth PP and FCP are trained with a population size of N = 32 agents which are sampled uniformly. For FCP, we use 3 checkpoints for each agent, therefore incurring no additional training burden: (1) at initialization (i.e. a low-skilled agent), (2) at the end of training (i.e. a fully-trained expert agent), and (3) at the middle of training, defined as when the agent reaches 50% of its final reward (i.e. an average-skilled agent). When varying architecture for the training partners of the FCP+A and\nFCP T,+A variants, we vary whether the partners use memory (i.e. LSTM vs not) and the width of their policy and value networks (i.e. 16 vs 256). In total, we train 8 agents for each of the 4 combinations, leaving the total population size of N = 32 unchanged, ensuring a fair comparison.\nTo train agents via behavioral cloning [58], we use the open-source Acme [30] to learn a policy from human gameplay data. Specifically, we collected 5 human-human trajectories of length 1200 time steps for each of the 5 layouts, resulting in 60k total environment steps. We divide this data in half and train two BC agents: (1) a partner for training a BCP agent, and (2) a “human proxy” partner for agent-agent evaluation. Following Carroll et al. [12], we use a set of feature-based observations for the agents (as opposed to RGB) and generate comparable results: performance is higher on 3 layouts (asymmetric, cramped, and ring) but poorer on the other 2 (circuit and forced).",
  "3 Related work": "Ad-hoc team play There is a large and diverse body of literature on ad-hoc team-play [5, 66], also known as zero-shot coordination [31]. Prior work based in game-theoretic settings has suggested the benefits of planning [71], online learning [51], and novel solution concepts [2], to name a few examples. More recently, multi-agent deep reinforcement learning has provided the tools to scale to more complex gridworld or continuous control settings, leading to work on hierarchical social planning [36], adapting to existing social conventions [40, 62], trajectory diversity [45], and theory of mind [14]. Ad-hoc team-play among novel agent partners is also an object of active study in the emergent communication literature [10, 11, 43]. This prior work has tended to focus on generalization to held-out agent partners as a proxy for human co-players.\nCollaborative play with novel humans has been evaluated more actively in the context of training agent assistants; see for instance [57, 68]. To our knowledge, our FCP agents represent the stateof-the-art in coordinating with novel human partners on an equal footing of capabilities in a rich gridworld environment, as measured by the challenge tasks in Carroll et al. [12].\nDiversity in multi-agent reinforcement learning In multi-agent reinforcement learning, agents that train with behaviorally diverse populations of game partners tend to demonstrate stronger performance than their self-play counterparts. For example, across a range of multi-agent games, generalization to held-out populations can be improved by training larger and more diverse populations [13, 42, 50]. In mixed-motive settings, cooperation among agents can be encouraged through social diversity, such as in player preferences and rewards [3, 47, 49]. Similarly, competitiveness can be optimized through selective matchmaking between increasingly diverse agents [24, 39, 69].\nDespite the increased focus on improving multi-agent performance, evaluation has typically been constrained to agent-agent settings. High-performing agents have infrequently been evaluated with humans, particularly in non-competitive domains [16]. We add to this growing literature, showing that training with diversity is a powerful approach for effective human-agent collaboration.\nHuman-agent interaction In recent years, increased attention has been directed toward designing machine learning agents capable of collaborating with humans [41, 57, 68, 72] (see also [16] for a broader review on Cooperative AI). Tylkin et al. [68] is particularly notable in also demonstrating that partially trained agents can be useful learning targets for human helpers, although in a different domain (cooperative Atari). Our method, FCP, can be seen as extending theirs by training with multiple “skill levels” and random seeds, rather than just one, which we demonstrate to be crucial to our agents’ performance (Tables 1 and 2 and Figure 7b).\nA key preceding entry in this research area is Carroll et al. [12], who similarly investigated humanagent coordination in Overcooked. We use their method (BCP) as a baseline throughout our experiments (Section 2.2). Relative to BCP, our approach removes the need for the expensive step of human data collection for agent training. Furthermore, through our novel human-agent experimental design, we go beyond objective performance metrics to compare the subjective preferences that agents generate. For a detailed comparison of methods and results, see Appendix E.",
  "4 Zero-shot coordination with agents": "In this section, we evaluate our FCP agent, its ablations, and the baselines with held-out agents.",
  "4.1 Evaluation method: collaborative evaluation with agent partners": "Our primary concern in this work is generalization to novel human partners (as investigated in Section 5). However, just as collecting human-human data for behavioral cloning is expensive, so too is evaluating agents with humans. Consequently, we instead use generalization to held-out agent partners as a cheap proxy of performance with humans. This is then used to guide our model selection process, allowing us to be more targeted with the agents we select for our human-agent evaluations.\nWe evaluate with three held-out populations:\n1. A BC model trained on human data, Hproxy, intended as a proxy of generalization to humans, as done by Carroll et al. [12].\n2. A set of self-play agents varying in seed, architecture, and training time (specifically, heldout seeds of the N = 32 partners trained for the FCP+A agent; see Section 2.4). These are intended to test generalization to a diverse yet still skillful population.\n3. Randomly initialized agents intended to test generalization to low-skill partners.\nFor all results, we report the average number of deliveries made by both players within an episode, aggregated across the 5 different layouts from Figure 4 (with the per-layout results reported in Appendix C.2). We estimate mean and standard deviation across 5 random seeds. For each seed, we evaluate the agent with all members of the held-out population for 10 episodes per agent-partner pair.",
  "4.2 Results": "Finding 1: FCP significantly outperforms all baselines To begin, we compare our FCP agent and the baselines when partnered with the three held-out populations introduced above. As can be seen in Figure 5, FCP significantly outperforms all baselines when partnered with all three held-out populations. Notably, it performs better than BCP with Hproxy, even though BCP trains with such a model and FCP does not. Similar to Carroll et al. [12], we find that BCP significantly outscores SP.\nWhen paired with a randomly initialized partner which behaves suboptimally, we see an even greater difference between FCP and the baselines. Given that FCP is trained with non-held-out versions of such agents, it may not be surprising that it does so well with partners that behave poorly. However, what is surprising is how brittle the other training methods are. This suggests that they may not perform well with humans who are not highly skilled players, which we will see in Section 5.\nFinding 2: Training with past checkpoints is the most beneficial variation for performance Next, we investigate how the different training partner variations influence FCP’s performance. In particular, we separately ablate the past checkpoints (T ) and architecture (A) variations, evaluating them with the same partners as in Figure 5. The results of this evaluation are presented in Table 1.\nComparing the FCP and FCP T columns, we see that removing past checkpoints from training significantly reduces performance. Comparing the FCP and FCP+A columns, we see that adding architectural variation to the training population offers no improvement over training with past\ncheckpoints. However, comparing the FCP T and FCP T,+A columns, we see that without training with past checkpoints, architectural variation in the population does improve performance.",
  "5 Zero-shot coordination with humans": "Ultimately, our goal is to develop agents capable of coordinating with novel human partners. In this section, we run an online study to evaluate our FCP agent and the baseline agents in collaborative play with human partners.",
  "5.1 Evaluation method: collaborative evaluation with human participants": "To test how effectively FCP’s performance generalizes to human partners, we recruited participants from Prolific [18, 55] for an online collaboration study (N = 114; 37.7% female, 59.6% male, 1.8% nonbinary; median age between 25–34 years). We used a within-participant design for the study: each participant played with a full cohort of agents (i.e. generated through every training method). This design allowed us to evaluate both objective performance as well as subjective preferences.\nParticipants first read game instructions and played a short tutorial episode guiding them through the dish preparation sequence (see Appendix D.1.1 for instruction text and study screenshots). Participants then played 20 episodes with a randomized sequence of agent partners and kitchen layouts. Episodes lasted T = 300 steps (1 minute) each. After every two episodes, participants reported their preference over the agent partners from those episodes on a five-point Likert-type scale. After playing all 20 episodes, participants completed a debrief questionnaire collecting standard demographic information and open-ended feedback on the study. Our statistical analysis below primarily relies upon the repeated-measures analysis of variance (ANOVA) method. See Appendix D for additional details of our study design and analysis, including independent ethical review.",
  "5.2 Results": "Finding 1: FCP coordinates best with humans, achieving the highest score across maps To begin, we compare the objective team performance supported by our FCP and baseline agents. The strong FCP performance observed in agent-agent play generalizes to human-agent collaboration:\nthe FCP-human teams significantly outperform all other agent-human teams, achieving the highest average scores across maps, every p < 0.001 (Figure 7a), while performing as well as or better than the other teams on each individual map (see Appendix D.3). Echoing the results from our agent-agent ablation experiments (Table 1), the inclusion of past checkpoints in training proves critical to FCP’s strong performance, p < 0.001 (Figure 7b). Similar to Carroll et al. [12], we find that BCP outscores SP when collaborating with human players, p < 0.001.\nFinding 2: Participants prefer FCP over all baselines FCP’s strong collaborative performance carries over to our participants’ subjective partner preferences. Participants expressed a significant preference for FCP partners over all other agents, including BCP, with every p < 0.05 (Figure 7c). Notably, while human-BCP and human-PP teams did not significantly differ in their completed deliveries, participants reported significantly preferring BCP over PP, p = 0.003, highlighting the informativeness of our subjective analysis.\n(a) Number of deliveries by partner (FCP and baselines). (b) Number of deliveries by partner (FCP and FCP T ). (c) Participant preference for row partner over column partner.\nFigure 7: Human-agent collaborative evaluation: Evaluation and preference metrics from humanagent play in episodes of length T = 300. Error bars represents 95% confidence intervals, calculated over episodes. Plots aggregate data across kitchen layouts; results calculated by individual layout can be found in Appendix D.3.",
  "5.3 Exploratory behavioral analysis": "To better understand how the human-agent scores and preferences may have arisen, here we analyze the resulting action trajectories of each human and agent player in our experiment.\nFinding 1: FCP exhibits the best movement coordination with humans First, we investigate how much each player moves in an episode (Figure 8a), where moving in a higher fraction of timesteps may suggest fewer collisions and thus better coordination with a partner. Notably, we observe two results: (1) humans rarely move, a behavior which is out-of-distribution for typical training methods (e.g. SP, PP) but is seen in the training distribution for BCP and FCP.\n(2) FCP moves the most on all layouts other than Forced, suggesting it is better at coordinating its movement strategy with its partner. This result was also reported by human participants, for example: “I noticed that some of my partners seemed to know they needed to move around me, while others seemed to get ‘stuck’ until I moved out of their way” (see Appendix D for more examples).\nFinding 2: FCP’s preferences over cooking pots aligns best with that of humans Next, we investigate whether there was a preference for a specific cooking pot in the layouts which included two cooking pots (Figure 8b). To do this, we calculate the difference in the number of times each pot was used by each player, where a high value indicates a strong preference for one pot and a low value indicates more equal preference for the two pots.\nAs can be seen in the FCP column, our agent typically has the most aligned preferences with that of humans (0.11 for FCP to 0.14 for humans). Behaviorally speaking, this means that our agent prefers one cooking pot over the other 55.5% of the time (i.e. a 0.11 point difference). In contrast, all other agents have a strong preference for a single pot. This is a non-adaptive strategy which generalizes poorly to typical human behavior of using both pots, leading to worse performance.",
  "6 Discussion": "Summary In this work, we investigated the challenging problem of zero-shot collaboration with humans without using human data in the training pipeline. To accomplish this, we introduced Fictitious Co-Play (FCP) – a surprisingly simple yet effective method based on creating a diverse set of training partners. We found that FCP agents scored significantly higher than all baselines when partnered with both novel agent and human partners. Furthermore, through a rigorous human-agent experimental design, we also found that humans reported a strong subjective preference to partnering with FCP agents over all baselines.\nLimitations and future work Our method currently relies on the manual process of initially training and selecting a diverse set of partners. This is not only time consuming, but also prone to researcher biases that may negatively influence the behavior of the created agents. Additionally, while we found FCP with a partner population size of N = 32 sufficient here, for more complex games, FCP may require an unrealistically large partner population size to represent sufficiently diverse strategies. To address these concerns, methods for automatically generating partner diversity for common-payoff games may be important. Possibilities include adaptive population matchmaking as been used in competitive zero-sum games [69], as well as auxiliary objectives that explicitly encourage behavioral diversity [19, 45, 46].\nOur method requires a known and fixed reward function. We also focus on one domain in order to compare with prior work which has argued that human-in-the-loop training is necessary. Consequently, the resulting agents are only designed to adaptively collaborate on a single task, and not to infer human preferences in general [1, 33, 59]. Moreover, if a task’s reward function is poorly aligned with how humans approach the task, our method may well produce subpar partners, as would any method without access to human data. Thus, additional domains and tasks should be studied to better understand how our method generalizes. Targeted experiments to test specific forms of generalization may be especially helpful in this regard [38], as could approaches that procedurally generate environment layouts requiring diverse solutions [22].\nFinally, it may be possible to produce even stronger agent assistants by combining the strengths of FCP (i.e. diversity) and BCP (i.e. human-like play). Indeed, Knott et al. [38] recently demonstrated that modifying BCP to train with multiple BC partners produces more robust collaboration with held-out agents, a finding that would be interesting to test with human partners.\nSocietal impact A challenge for this line of work is ensuring agent behavior is aligned with human values (i.e. the AI value alignment problem [23, 59]). Our method has no guarantees that the resulting policy aligns with the preferences, intentions, or welfare of its potential partners. It likewise does not exclude the possibility that the target being optimized for is harmful (e.g. if the agent’s partner expresses preferences or intentions to harm others). This could therefore produce negative societal effects either if training leads to poor alignment or if agents are optimized for harmful metrics.\nOne potential strategy for mitigating these risks is the use of human preference data [15]. Such data could be used to fine-tune and filter trained agents before deployment, encouraging better alignment with human values. A key question in this line of research is how human preference data should be\naggregated—or selected, in the case of expert preferences—when our aim is to create socially aligned agents (i.e. agents that are sufficiently aligned for everyone). Relatedly, targeted research on human beliefs and perceptions of AI [48], and how they steer human-agent interaction, would help inform agent design for positive societal impact. For instance, developers could incorporate specific priors into agents to reinforce tendencies for fair outcomes [20, 32].\nConclusion We proposed a method which is both effective at collaborating with humans and simple to implement. We also presented a rigorous and general methodology for evaluating with humans and eliciting their preferences. Together, these establish a strong foundation for future research on the important challenge of human-agent collaboration for benefiting society.",
  "Acknowledgements": "The authors would like to thank Mary Cassin for creating the game sprite art; Rohin Shah, Thore Graepel, and Iason Gabriel for feedback on the draft; Lucy Campbell-Gillingham, Tina Zhu, and Saffron Huang for support in evaluating agents with humans; and Max Kleiman-Weiner, Natasha Jaques, Marc Lanctot, Mike Bowling, and Dan Roberts for useful discussions.\nFunding disclosure\nThis work was funded solely by DeepMind. The authors declare no competing interests.",
  "Reviewer Summary": "Reviewer_1: This paper proposes a method of training collaborator agents that cooperates with humans without the need for human behaviour data. It proposes that the main drawback of \"behaviour cloning play\"(BCP) is that it is expensive to collect sufficient human data. The method proposed in this paper is called Fictitious Co-Play(FCP), and is based on the prior methods self-play(SP). Through different settings, initialization and additional checkpoints, we can get a collection of self-play agents of different degrees of skill. Via training with the collection of self-play agents, a collaborator agent which collaborates with novel human well in common-payoff games can be obtained. Self-play agents and their checkpoints during training replace a large amount of human data.\n\nReviewer_2: This paper seeks to build agents that can effectively collaborate with humans when playing \"common payoff games.\" They propose to modify Fictitious Self-Play (FSP) methods to work in this setting, hence the name Fictitious Co-Play (FCP). Specifically, a final agent is trained to collaborate with a diverse range of agents (generated w/ varying random seed) alongside earlier checkpoints of those agents. This simulates a self-play pool with varying skill levels and varying strategies. Thanks to this more diverse training experience, the final FCP agent is more robust to unseen strategies and more amenable to human interaction.\n\nReviewer_3: This paper proposes a model-free RL training method to improve robustness to partner/teammate changes, with a focus on human partners. The key finding is that diversity in behavior appears sufficient; as long as the agent has been exposed to similar partner behaviors during training, performance is good. The authors propose to incorporate such behaviors by using agents trained using different random seeds or from previous checkpoints in the training process. Experiments in the Overcooked domain suggest this fictitious co-play (FCP) setup enables agents to better cooperate with both artificial and human agents."
}