Artist: Open AI
What we can learn from Open AI 5

Imitating the bots is a flawed concept because the AI sees and thinks about the game differently and because it operates under different rules than normal Dota. Nonetheless, this doesn’t mean we can’t learn from it if we’re smart!

In the past, the 1v1 Bot actually shifted the mid meta successfully. It was using Mangos and Salves much more frequently than a human would, which allowed it to fight for harass and farm with no downtime and snowball into a lane win (the AI prioritized Mangos and Salves above not only other consumables but items like Wraith Bands, Boots, etc.).

This strategy is optimal for 1v1 SF mid, yet it doesn’t account for the fact that after the laning stage you need actual items to win the game. Because of this, mid players integrate this into their gameplay, yet didn’t take it to such an extreme. Salves and Mangos are great tools to pressure your lane opponent and win the lane, so nowadays they are used more commonly than Tangoes and Clarities mid, yet over-doing to the point it hurts your item progression isn’t necessary. I.e. you can still learn from the bot without imitating it.

Open AI 5 is a bit different than the 1v1 bot. As explained in more detail in this brilliant Medium post linked above, the bots receive short-term rewards that help them work toward their long-term goal (winning). You can check out the reward list here. Understanding those rewards is pretty important to put the bot’s decision making into context and not to draw wrong conclusions.

1. Rage Buybacks

Many people noticed that the bots buyback instantly a lot of the time even in the early game when no objectives are under threat. This lead to speculations that instantly buying back might be beneficial because of multiple reasons:

  • You don’t lose that much gold early in the game
  • It helps you instantly push-out lanes and recover map control
  • You’re effectively trading gold for XP

Yet, although those reasons have a grain of truth to them (nothing is black or white, every in-game decision involves some kind of tradeoff), the reason why the bots are doing it is most likely not that deep:

[1]: The agent receives reward when it gains gold but does not lose reward when it loses gold (e.g. by buying an item.)

This likely means that it doesn’t lose reward when buying back as well, yet it gains points for farming a little bit for the time that it is not dead after buying back, which makes the action net-positive in terms of points. The bot doesn’t have a long-term strategy for winning the game (it chases the short-term rewards and it stumbles into the ancient kill, which is just another reward to it), and the fact that such buybacks hurt its long-term economy and chances to win the game are probably ignored.

LESSON 1:

Even though the bots buy back instantly almost all the time, you probably should NOT follow this example because of the different rules under which the bots operate.

2. Equal Farm Distribution

The other big conclusion that a lot of people made is that an even net-worth distribution is better than the current standard farm distribution which includes an over-farmed pos. 1 and 2 and a severely under-farmed pos. 5 (with pos. 3 and 4 hovering between greedy and sacrificial mostly based on heroes).

This, however, is once more not a safe conclusion to make because of the rules under which the bots are operating. Remember that as a human player the only thing you ultimately care about is to kill the enemy Ancient before they kill yours. Going 0-10-0 and winning 10 straight games is much better than going 10-0-10 and losing 10 games.

This is not true for the bots:

  • One, they value every single reward they are given for the length of the game (last hits, gold, and XP, kills, deaths, etc.).
  • Two, no single bot thinks of himself as support or carry. All of them have an equal reward for last-hits, gold, and XP. By farming (even as support) you are contributing towards the team total score (the “team-spirit” value makes sure the bots value the team total more than their individual score). More importantly, however, the bot’s don’t have a long time horizon for their decision making. Their long-term strats are basically up to 5 minutes in the future.

The sacrificial support “strategy” is born from the simple concept that the long-term impact of some heroes is strongly correlated to their levels of farm, while other heroes have a similar impact regardless of items. I.e. some heroes use gold better than others. Sacrificing all of the farm on your Oracle for the first 30 minutes of the game in order to allow your Spectre to destroy the enemy team in the last 10 minutes makes sense to humans.

It doesn’t make sense to the bots – they cannot plan for the last 10 minutes of the game from the very beginning, and in that sense the “support and carry” concepts are lost to them. It nets better reward scores for them to simply farm with all heroes to the best of their ability.

LESSON 2:

That being said, this doesn’t mean that there is no benefit to supports with farm. Teams like Fnatic, Secret, old EG with Zai, and even TI3 Alliance, showcased that having supports with more items than your opponents gives you a big advantage in fights. The question is what you sacrifice to get this farm. If you are sacrificing map control and the farm of the heroes on your team who really need it in the long run, playing too greedy as a support will still bring you negative results on average. Farming the jungle as CM with right-clicks is probably not worth it – it is very likely that there are other things to do that will bring you more benefits in the long run.

3. Deathball – best strat?

The tactics of the bots are pretty simple.

  1. They try to win the lanes.
  2. They farm on all heroes.
  3. They group up as five, take favorable fights and objectives.

There are three reasons this makes a lot of sense for them.

One, they have 10,000 years of experience playing 17 heroes, which helps them judge the outcome of an engagement extremely accurately. This means if the bots are trying to force a fight, they have mathematical evidence they are going to win it.

Two, they don’t communicate as humans do. All of them see the same opportunities all over the map and make the same individual decisions. This helps them instantly jump on an opportunity as a unit. Humans take time for such team-moves. Someone has to see the opportunity, communicate it to their teammates with words (or at least pings if it’s a simple move), and then act on it. This makes jumping on favorable and avoiding unfavorable fights harder.

Three, the best counter to 5-man Dota is Rat Dota – split pushing, cutting the waves, avoiding said unfavorable fights. Yet, the 17 hero pool doesn’t have some of the best Rat heroes, which means you don’t have the paper in the rock-paper-scissors game, which means Rock is always the best strat.

LESSON 3:

Bots are definitely over-valuing 5-man Dota because of the restricted hero pool and their “inhuman” capabilities. Yet, in my experience pubs are doing the opposite – they are under-valuing 5-man Dota. In terms of team coordination, it’s the simplest strategy to execute, and the counter – Rat Dota, is probably the hardest. This means picking 5-man heroes and trying to convince your teammates to stick and fight together will on average yield positive results for you.

4. Kills - Overrated?

Below you can see some of the rewards the Open AI devs created for the bots  -the breadcrumbs that lead them to victory:

Individual:                    Weight:                Awarded for:

  • Experience         0.002                    Per unit of experience.
  • Gold                      0.006                    Per unit of gold gained[1].
  • Mana                    0.75                      Mana (fraction of total).
  • Hero Health        2.0                        Gaining (or losing) health[2].
  • Last Hit                0.16                      Last Hitting an enemy creep[3].
  • Deny                     0.2                        Last Hitting an allied creep[3].
  • Kill                        -0.6                       Killing an enemy hero[3].
  • Death                   -1.0                        Dying.

One thing immediately pops to mind. A kill has a negative weight. The following explanation is included:

[3]: This score supplements the score for the gold/experience gained. The explicit "Kill" score is negative to reduce the agents' reward received by a kill, but the total is still positive.

So, Kills still bring a positive reward to the bots, but the devs deemed it’s necessary to deter them from going for kills too often. This is quite curious, especially bearing in mind the bots should be able to judge risk better than humans and don’t have an ego which tells them they are awesome when they kill other bots.

The over-aggression of the bots might be due to them not planning for the long run strategically and not being punished for losing gold (yet being rewarded for gaining it), yet the death itself has a negative value, so this should balance things out. Whatever the case, it’s safe to assume that the devs played with the numbers, and the values quoted above gave the best results for the experiment.

LESSON 4:

This leads us to the most prosaic and oldest lesson in Dota – kills are overrated. Yes, they are nice but prioritizing map control, your own resource development, and objectives give better results in the long run. Farming instead of taking risky fights is probably the No. 1 lesson a low-level player needs to internalize in order to start winning more often.


It would be extremely cool to see the bots play without the item and hero restrictions to see how they adapt and what meta they form. Sadly, Open AI mentioned that their ultimate goal is not to make the best Dota bot, which makes the resources required to include the whole hero pool unnecessary. The Dota bot is simply a proof of concept that they can make a general purpose bot that can learn and excel in complicated environments and tasks.

(In other words, after beating you at your hobby, the bot wants to beat you at your job as well! Yay!)

I hope you found the lessons above interesting! If you did, we have similar articles about pro games!